Clinical Validation of Artificial Intelligence-Powered PD-L1 Tumor Proportion Score Interpretation for Immune Checkpoint Inhibitor Response Prediction in Non-Small Cell Lung Cancer

PURPOSE

Evaluation of PD-L1 tumor proportion score (TPS) by pathologists has been very impactful but is limited by factors such as intraobserver/interobserver bias and intratumor heterogeneity. We developed an artificial intelligence (AI)-powered analyzer to assess TPS for the prediction of immune checkpoint inhibitor (ICI) response in advanced non-small cell lung cancer (NSCLC).

MATERIALS AND METHODS

The AI analyzer was trained with 393,565 tumor cells annotated by board-certified pathologists for PD-L1 expression in 802 whole-slide images (WSIs) stained by 22C3 pharmDx immunohistochemistry. The clinical performance of the analyzer was validated in an external cohort of 430 WSIs from patients with NSCLC. Three pathologists performed annotations of this external cohort, and their consensus TPS was compared with AI-based TPS.

RESULTS

In comparing PD-L1 TPS assessed by AI analyzer and by pathologists, a significant positive correlation was observed (Spearman coefficient = 0.925; P < .001). The concordance of TPS between AI analyzer and pathologists according to TPS ≥50%, 1%-49%, and <1% was 85.7%, 89.3%, and 52.4%, respectively. In median progression-free survival (PFS), AI-based TPS predicted prognosis in the TPS 1%-49% or TPS <1% group better than the pathologist's reading, with the TPS ≥50% group as a reference (hazard ratio [HR], 1.49 [95% CI, 1.19 to 1.86] v HR, 1.36 [95% CI, 1.08 to 1.71] for TPS 1%-49% group, and HR, 2.38 [95% CI, 1.69 to 3.35] v HR, 1.62 [95% CI, 1.23 to 2.13] for TPS <1% group).

CONCLUSION

PD-L1 TPS assessed by AI analyzer correlates with that of pathologists, with clinical performance also being comparable when referenced to PFS. The AI model can accurately predict tumor response and PFS of ICI in advanced NSCLC via assessment of PD-L1 TPS.

Tumor-infiltrating lymphocyte enrichment predicted by CT radiomics analysis is associated with clinical outcomes of non-small cell lung cancer patients receiving immune checkpoint inhibitors

Background

Enrichment of tumor-infiltrating lymphocytes (TIL) in the tumor microenvironment (TME) is a reliable biomarker of immune checkpoint inhibitors (ICI) in non-small cell lung cancer (NSCLC). Phenotyping through computed tomography (CT) radiomics has the overcome the limitations of tissue-based assessment, including for TIL analysis. Here, we assess TIL enrichment objectively using an artificial intelligence-powered TIL analysis in hematoxylin and eosin (H&E) image and analyze its association with quantitative radiomic features (RFs). Clinical significance of the selected RFs is then validated in the independent NSCLC patients who received ICI.

Methods

In the training cohort containing both tumor tissue samples and corresponding CT images obtained within 1 month, we extracted 86 RFs from the CT images. The TIL enrichment score (TILes) was defined as the fraction of tissue area with high intra-tumoral or stromal TIL density divided by the whole TME area, as measured on an H&E slide. From the corresponding CT images, the least absolute shrinkage and selection operator model was then developed using features that were significantly associated with TIL enrichment. The CT model was applied to CT images from the validation cohort, which included NSCLC patients who received ICI monotherapy.

Results

A total of 220 NSCLC samples were included in the training cohort. After filtering the RFs, two features, gray level variance (coefficient 1.71 x 10^-3) and large area low gray level emphasis (coefficient -2.48 x 10^-5), were included in the model. The two features were both computed from the size-zone matrix, which has strength in reflecting intralesional texture heterogeneity. In the validation cohort, the patients with high predicted TILes (≥ median) had significantly prolonged progression-free survival compared to those with low predicted TILes (median 4.0 months [95% CI 2.2-5.7] versus 2.1 months [95% CI 1.6-3.1], p = 0.002). Patients who experienced a response to ICI or stable disease with ICI had higher predicted TILes compared with the patients who experienced progressive disease as the best response (p = 0.001, p = 0.036, respectively). Predicted TILes was significantly associated with progression-free survival independent of PD-L1 status.

Conclusions

In this CT radiomics model, predicted TILes was significantly associated with ICI outcomes in NSCLC patients. Analyzing TME through radiomics may overcome the limitations of tissue-based analysis and assist clinical decisions regarding ICI.

Abstract 6172: Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals immune-excluded phenotype is correlated with TGF-beta pathway related genomic aberrations

Aberrant transforming growth factor-beta(TGF-B) pathway in the tumor microenvironment has been highlighted as one of the core resistance pathways of immunotherapy, by excluding tumor-infiltrating lymphocytes (TIL) out of the tumor area. However, no studies have coupled immune phenotypes classified by spatial analysis of TIL in whole slide images (WSI) with TGF-B pathway analysis on a large-scale database. Here, we hypothesized that the immune-excluded phenotype classified by a deep-learning spatial analysis model, Lunit SCOPE IO, would be correlated with the aberrant TGF-B pathway in The Cancer Genome Atlas (TCGA) cohorts. Aberrant TGF-B pathway was measured by Trimmed Mean of M-values (TMM) normalized and transformed to log2 of counts-per-million of previously published gene sets of fibroblast-specific TGF-beta responsive gene signature, using edgeR packages from TCGA RNA-sequencing data (n=6,709) across the 23 cancer types. Lunit SCOPE IO was developed to identify immune phenotypes trained and validated from 3,166 multi-cancer H&E WSI with sections of 2.8e+9 mm2 tumor tissue containing 5.9e+6 TIL annotated by 52 board-certified pathologists. Lunit SCOPE IO classified immune phenotypes as immune-inflamed and -excluded according to the proportion of TIL density either highly conserved in cancer epithelium (CE) or cancer stroma (CS), respectively, and otherwise, classified as immune-desert with low TIL density in CE and CS. Aberrant TGF-B expression was highly enriched in multiple cancer types including pancreatic cancer, head and neck cancer, kidney clear cell carcinoma, lung squamous cell carcinoma, and breast cancer, in ascending order. TGF-B expression was increased in microsatellite-stable tumor samples (p = 7.4e-15) or samples with low tumor mutational burden (TMB, &lt; 10/megabase, p = 4.9e-8), compared to those with microsatellite instability-high or high TMB, respectively. Interestingly, TGF-B expression was highly correlated with the proportion of cancer stroma in WSI (R = 0.315, p &lt; 2.2e-16) and the proportion of immune-excluded phenotype (R = 0.115, p &lt; 2.2e-16) across multiple cancer types. Tumor samples with SMAD4 mutations (n = 161, 2.4%) had significantly higher TGF-B expression (p = 0.0190), and a higher proportion of immune-excluded phenotype (p &lt; 2.2e-16) in WSI, compared to wild-type SMAD4. Aberrant TGF-B pathway is clearly associated with increased proportion of cancer stroma, and excluded TIL, or immune-excluded phenotype in a large-scale pan-carcinoma analysis.

Citation Format: Gahee Park, Sanghoon Song, Hyung-Gyo Cho, Soo Ick Cho, Wonkyung Jung, Lunit AI team, Sergio Pereira, Donggeun Yoo, Kyunghyun Paeng, Chan-Young Ock. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals immune-excluded phenotype is correlated with TGF-beta pathway related genomic aberrations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 6172.

Abstract 644: Deep learning-based H&E analyzer reveals distinct immune profiles and clinical outcomes among immune phenotypes in uterine corpus endometrial carcinoma

Introduction: Deep learning-based H&E analyzer can classify the tumor microenvironment as three immune phenotypes: the immune-inflamed, excluded and desert. Our previous study demonstrated a distinct transcriptomic and immunologic landscape amongst the phenotypes in non-small cell lung cancer (NSCLC). However, it has not been fully investigated in other cancers. Here, we explore the immune profiles and clinical outcomes between the three immune phenotypes in uterine corpus endometrial carcinoma (UCEC).

Methods: Tissue H&E slide images, sequencing data, and clinical data were utilized from The Cancer Genome Atlas (TCGA). Lunit-SCOPE IO was trained with multi-cancer 3,166 H&E whole slide images annotated by pathologists. Based on the proportion of tumor infiltrating lymphocytes (TIL) highly conserved either in cancer epithelium (CE) or cancer stroma (CS), Lunit-SCOPE IO classifies tumors as immune-inflamed and excluded, respectively. Also, it classifies tumors with low TIL density in CE and CS as immune-desert.

Results: Among 486 patients with UCEC, the frequency of immune-inflamed, excluded and desert was 174 (35.8%), 160 (32.9%), and 156 (32%), respectively. In the three subgroup comparison, immune-inflamed was associated with the best survival outcome and -excluded was associated with the worst survival outcome (Inflamed vs excluded, HR 0.30 95% CI 0.17-0.55, p&lt;.001; desert vs excluded, HR 0.50 95% CI 0.30-0.84, p=0.009). Likewise, inflamed subtype showed better overall survival (HR 0.43, 95% CI 0.25-0.75, p=0.003) compared to others. In microsatellite instability high (MSI-H) tumors, we observed a similar tendency of improved overall survival in the tumors of inflamed subtype, both compared to the excluded subtype and to a combination of other subtypes. (Inflamed vs excluded, HR 0.18 95% CI 0.05-0.73, p=0.017; inflamed vs others, HR 0.21 95% CI 0.06-0.72, p=0.014). Immune-inflamed had significantly higher cytolytic activity (Inflamed 7.25 vs others 6.34, p&lt;.001) and was associated with higher PD-L1 expression (Inflamed 19.03 vs others 10.7, p=0.003) and CTLA4 expression (Inflamed 60.62 vs others 31.5, p&lt;.001). Immune-inflamed had a higher proportion of CD8 positive T cell (Inflamed 16.7% vs 12.8%, p&lt;.001) and M1 macrophage (Inflamed 3.9% vs others 2.8%, p&lt;.001) and a lower proportion of M2 macrophage (Inflamed 15% vs others 17.9%, p&lt;.001).

Conclusion: The three tissue phenomic subtypes showed distinct immune profiles and clinical outcomes, with immune-inflamed having the best overall survival outcome. In particular, non-inflamed group was associated with worse overall survival even in MSI-H tumors deemed to have more favorable prognosis compared to MSS tumors. Given the definite differences in the survival outcome, tissue H&E based tumor microenvironment classification may serve as a potential prognostic biomarker in UCEC.

Citation Format: Horyun Choi, Leeseul Kim, Jinah Kim, Yeun Ho Lee, Hyung-Gyo Cho, Na Hyun Kim, Gahyun Gim, Sanghoon Song, Gahee Park, Soo Ick Cho, Sergio Pereira, Donggeun Yoo, Kyunghyun Paeng, Chan-Young Ock, Young Kwang Chae. Deep learning-based H&E analyzer reveals distinct immune profiles and clinical outcomes among immune phenotypes in uterine corpus endometrial carcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 644.

Artificial intelligence (AI)–powered spatial analysis of tumor-infiltrating lymphocytes (TIL) for prediction of response to neoadjuvant chemotherapy (NAC) in triple-negative breast cancer (TNBC)

595

Background: Stromal TIL are a well-recognized prognostic and predictive biomarker in breast cancer. There is a need for tools assisting visual assessment of TIL, to improve reproducibility as well as for convenience. This study aims to assess the clinical significance of AI-powered spatial TIL analysis in the prediction of pathologic complete response (pCR) after NAC in TNBC patients. Methods: H&E stained slides and clinical outcomes data were obtained from stage I – III TNBC patients treated with NAC in two centers in Korea. For spatial TIL analysis, we used Lunit SCOPE IO, an AI-powered H&E Whole-Slide Image (WSI) analyzer, which identifies and quantifies TIL within the cancer or stroma area. Lunit SCOPE IO was developed with a 13.5 x 109 micrometer2 area and 6.2 x 106 TIL from 17,849 H&E WSI of multiple cancer types, annotated by 104 board-certified pathologists. iTIL score and sTIL score were defined as area occupied by TIL in the intratumoral area (%) and the surrounding stroma (%), respectively. Immune phenotype (IP) of each slide was defined from spatial TIL calculation, as inflamed (high TIL density in tumor area), immune-excluded (high TIL density in stroma), or desert (low TIL density overall). Results: A total of 954 TNBC patients treated from 2006 to 2019 were included in this analysis. pCR (ypT0N0) was confirmed in 261 (27.4%) patients. The neoadjuvant regimens used were mostly anthracycline (97.8%) and taxane (75.1%) -based, with 116 (12.1%) patients receiving additional platinum and 41 (4.3%) patients treated as part of immune checkpoint inhibitor or PARP inhibitor clinical trials. The median iTIL score and sTIL score were 4.3% (IQR 3.2 – 5.8) and 8.1% (IQR 6.3 – 13.4), respectively. The mean iTIL score was significantly higher in patients who achieved pCR after NAC (5.8% vs. 4.5%, p < 0.001), and a similar difference was observed with sTIL score (12.1%.1 vs. 9.4%, p < 0.001). iTIL score was found to remain as an independent predictor of pCR along with cT stage and Ki-67 in the multivariable analysis (adjusted odds ratio 1.211 (95% CI 1.125 – 1.304) per 1 point (%) change in the score, p <0.001). By IP groups, 291 (30.5%) patients were classified as inflamed, 502 (52.6%) as excluded, and 161 (16.9%) as desert phenotype. The patients with inflamed phenotype were more likely to achieve pCR (44.7%) than other phenotypes (19.8%, p < 0.001). Conclusions: AI-powered spatial TIL analysis could assess TIL densities in the cancer area and surrounding stroma of TNBC, and TIL density scores and IP classification could predict pCR after NAC.

Artificial intelligence-powered human epidermal growth factor receptor 2 (HER2) analyzer in breast cancer as an assistance tool for pathologists to reduce interobserver variation

e12543

Background: Human epidermal growth factor receptor 2 (HER2) expression is a predictive marker for HER2-targeted therapy in breast cancer patients. Interobserver variation in the interpretation of HER2 levels exists among pathologists, thus a method to increase the consistency of evaluation is needed. This study aimed to evaluate the performance of the artificial intelligence (AI)-based Lunit SCOPE HER2 in assisting pathologists to evaluate HER2 expression levels in breast cancer. Methods: Lunit SCOPE HER2 was developed with a 1.04 x 1010 μm2 area and 7.31 x 105 tumor cells from 1,133 HER2 immunohistochemistry stained whole-slide images (WSI) of breast cancer, annotated by 113 board-certified pathologists. The AI model was developed based on a semantic segmentation algorithm, which consists of two atrous spatial pyramid pooling blocks for tissue level classification and for tumor cell level classification. To validate the model, a total of 209 HER2 WSIs diagnosed with breast cancer were obtained from Kyung Hee University Hospital in Korea and were assigned as an external validation set. Three board-certified pathologists evaluated slide level HER2 expression (3+, 2+, 1+, and 0) twice, first without AI assistance and second, with it. The second reading was performed for WSIs where the pathologist's reading showed discrepancy with the AI model. Results: In the external validation set, all pathologists scored the same HER2 grade in 103 WSIs (49.3%), and the Fleiss kappa value was 0.512. The HER2 grade from the AI model and pathologists was the same in 151 WSIs (72.2%), and the weighted kappa value was 0.844. The pathologists re-evaluate 43, 63, and 83 WSIs, respectively. After AI assistance, all pathologists scored the same HER2 grade in 156 WSIs (74.6%), and the Fleiss kappa value increased to 0.762 (Table). Conclusions: This study demonstrates that an AI-powered HER2 analyzer can help achieve consistent HER2 expression level evaluation in breast cancer by reducing interobserver variability. Thus, the AI model can be applied as an assistance tool for pathologists in HER2 grade evaluation.[Table: see text]

Artificial intelligence-powered whole-slide image analyzer reveals a distinctive distribution of tumor-infiltrating lymphocytes in neuroendocrine tumors and carcinomas

e16214

Background: Immune checkpoint inhibitors (ICIs) have shown promising treatment outcomes for various types of tumors. However, in neuroendocrine tumors and carcinomas (NET/NEC), ICI has proven to be applicable for only limited cases. In addition, little is known about the immunoprofile of NET/NEC. Here we investigate the landscape of tumor-infiltrating lymphocytes (TIL) using artificial intelligence (AI)-powered H&E whole-slide image (WSI) analyzer to elucidate the tumor microenvironment of NET/NEC. Methods: A total of 240 H&E stained pathologic slides diagnosed with NET/NEC were obtained from Ajou University Medical Center in Korea (from January 2020 to December 2021). For spatial TIL analysis, we used Lunit SCOPE IO, an AI-powered H&E WSI analyzer, which identifies and quantifies TIL within the cancer or stroma area. The AI was developed with a 13.5 x 109 μm2 area and 6.2 x 106 TIL from 17,849 H&E WSI of multiple cancer types, annotated by 104 board-certified pathologists. Intra-tumoral TIL, stromal TIL, and combined (cancer + stroma) TIL density were defined as the TIL count divided by the area of interest respectively. NET with histological grade 1 and 2 were labeled as low grade and NET with histological grade 3 and together with NEC were labeled as high grade. Primary origins of the NET/NEC were grouped by colorectum, stomach, small intestine, hepatopancreatobiliary, lung, and other organs (including anus, appendix, breast, cervix, and larynx). Results: Total slides classified as low grade and high grade were 211 and 29, respectively; 175 samples were from colorectal, 19 from stomach, 16 from small intestine, 16 from hepatopancreaticobiliary, seven from lung, and seven from other organs. The median intra-tumoral TIL, stromal TIL, and combined TIL density were 4.2/mm2 (IQR 1.718 - 11.478), 139.1/mm2 (IQR 75.4 - 313.9), and 62.4/mm2 (IQR 36.3 - 162.6), respectively. The median intra-tumoral TIL density was significantly higher in patients with high grade NET/NEC compared with low grade (11.9/mm2 [IQR 4.51 - 30.9] vs 3.45/mm2 [IQR 1.63 - 9.81], p < 0.001). However, statistical differences in stromal TIL density and combined TIL density were not observed between low grade and high grade NET/NEC. The highest intra-tumoral TIL density in the group classified according to primary origins was lung (n = 7, median: 16.5/mm2, IQR 5.01 - 34.1) and was followed by stomach (n = 19, median: 11.8/mm2, IQR 8.64 - 20.8), and small intestine (n = 16, median: 7.23/mm2, IQR 4.12 - 25.2). Conclusions: AI-powered TIL analysis reveals that the intra-tumoral TIL density is significantly higher in high grade NET/NEC than low grade NET. Our findings align with recent evidence that ICIs are effective against large cell NEC and small cell carcinoma.Therefore, AI-powered TIL analysis should be investigated as a predictive biomarker for ICI response in NET/NEC.

Tumor-infiltrating lymphocyte enrichment predicted by CT radiomic analysis is associated with clinical outcomes of immune checkpoint inhibitor in non–small cell lung cancer

2663

Background: Enrichment of tumor-infiltrating lymphocytes (TIL) in the tumor microenvironment (TME) is a reliable biomarker of immune checkpoint inhibitors (ICI) in non-small cell lung cancer (NSCLC). Phenotyping through CT radiomics has the overcome the limitations of tissue-based assessment, including for TIL analysis. Here, we objectively assess TIL enrichment using an artificial intelligence-powered H&E analyzer, Lunit SCOPE IO, and analyze its association with advanced quantitative imaging features extracted via radiomic analysis. Clinical significance of the selected radiomic features (RFs) is then validated in independent NSCLC patients who received ICI. Methods: In the training cohort, which included 235 NSCLC patients with both tumor tissue and corresponding CT images obtained within 1 month, we extracted 86 RFs from the CT images. From tissue, a patient’s TIL enrichment score (TILes) was defined as the fraction of tissue area with high intra-tumoral or stromal TIL density, divided by the whole TME area, as measured on an H&E slide. From the corresponding CT images, the least absolute shrinkage and selection operator model was then developed using features that were significantly associated with TIL enrichment. The CT model was then applied to CT images from the validation cohort, which included 242 NSCLC patients who received ICI as ≥ second line. Results: Among the extracted RFs, 22 features were significantly associated with TILes (p < 0.005). After excluding features of multicollinearity and/or zero-coefficient, two features, gray level variance (coefficient 1.71 x 10-3) and low gray level emphasis (coefficient -2.48 x 10-5), were finally included in the model. The two features were both computed from the size-zone matrix (SZM), the idea of which is to break down a given tumor volume into smaller spatially contiguous compartments of different sizes. In the validation cohort, the patients with high predicted TILes (≥ median) had significantly prolonged progression-free survival compared with those with low predicted TILes (median 3.81 months [95% CI 2.14 – 5.69] versus 1.94 months [95% CI 1.58 – 2.93], hazard ratio 0.69 [95% CI 0.53 – 0.90], p = 0.007). Conclusions: This CT radiomics model is able to assess TIL enrichment in TME, which is significantly associated with favorable ICI outcomes in NSCLC. Analyzing the TME through radiomics may overcome limitations of tissue-based analysis and inform clinical decisions, particularly related to use of ICI.

Artificial Intelligence-Powered Hematoxylin and Eosin Analyzer Reveals Distinct Immunologic and Mutational Profiles among Immune Phenotypes in Non-Small-Cell Lung Cancer

The tumor microenvironment can be classified into three immune phenotypes: inflamed, immune excluded, and immune-desert. Immunotherapy efficacy has been shown to vary by phenotype; yet, the mechanisms are poorly understood and demand further investigation. This study unveils the mechanisms using an artificial intelligence-powered software called Lunit SCOPE. Artificial intelligence was used to classify 965 samples of non-small-cell lung carcinoma from The Cancer Genome Atlas into the three immune phenotypes. The immune and mutational profiles that shape each phenotype using xCell, gene set enrichment analysis with RNA-sequencing data, and cBioportal were described. In the inflamed subtype, which showed higher cytolytic score, the enriched pathways were generally associated with immune response and immune-related cell types were highly expressed. In the immune excluded subtype, enriched glycolysis, fatty acid, and cholesterol metabolism pathways were observed. The KRAS mutation, BRAF mutation, and MET splicing variant were mostly observed in the inflamed subtype. The two prominent mutations found in the immune excluded subtype were EGFR and PIK3CA mutations. This study is the first to report the distinct immunologic and mutational landscapes of immune phenotypes, and demonstrates the biological relevance of the classification. In light of these findings, the study offers insights into potential treatment options tailored to each immune phenotype.

Artificial Intelligence-Powered Spatial Analysis of Tumor-Infiltrating Lymphocytes as Complementary Biomarker for Immune Checkpoint Inhibition in Non-Small-Cell Lung Cancer

PURPOSE

Biomarkers on the basis of tumor-infiltrating lymphocytes (TIL) are potentially valuable in predicting the effectiveness of immune checkpoint inhibitors (ICI). However, clinical application remains challenging because of methodologic limitations and laborious process involved in spatial analysis of TIL distribution in whole-slide images (WSI).

METHODS

We have developed an artificial intelligence (AI)–powered WSI analyzer of TIL in the tumor microenvironment that can define three immune phenotypes (IPs): inflamed, immune-excluded, and immune-desert. These IPs were correlated with tumor response to ICI and survival in two independent cohorts of patients with advanced non–small-cell lung cancer (NSCLC).

RESULTS

Inflamed IP correlated with enrichment in local immune cytolytic activity, higher response rate, and prolonged progression-free survival compared with patients with immune-excluded or immune-desert phenotypes. At the WSI level, there was significant positive correlation between tumor proportion score (TPS) as determined by the AI model and control TPS analyzed by pathologists (P < .001). Overall, 44.0% of tumors were inflamed, 37.1% were immune-excluded, and 18.9% were immune-desert. Incidence of inflamed IP in patients with programmed death ligand-1 TPS at < 1%, 1%-49%, and ≥ 50% was 31.7%, 42.5%, and 56.8%, respectively. Median progression-free survival and overall survival were, respectively, 4.1 months and 24.8 months with inflamed IP, 2.2 months and 14.0 months with immune-excluded IP, and 2.4 months and 10.6 months with immune-desert IP.

CONCLUSION

The AI-powered spatial analysis of TIL correlated with tumor response and progression-free survival of ICI in advanced NSCLC. This is potentially a supplementary biomarker to TPS as determined by a pathologist.

Abstract P4-05-07: Assistance with an artificial intelligence-powered tumor infiltrating lymphocytes (TIL) analyzer reduces interobserver variation in pathologic scoring of TIL in breast cancer

Background Tumor infiltrating lymphocytes (TIL) is a promising prognostic marker in breast cancer. However, TIL is manually scored by pathologists, thus laborious work is required and interobserver heterogeneity exists in the results. In this study, we aimed to evaluate the clinical utility of an artificial intelligence (AI)-powered TIL analyzer in terms of reducing the interobserver variation. Methods Lunit SCOPE IO, AI-powered TIL analyzer was trained and validated with a 2.8 x 109 micrometer2 area and 5.9 x 106 TIL from 3,166 H&E Whole-Slide Images (WSI) of multiple cancer types including breast cancer, annotated by 52 board-certified pathologists. Three independent board-certified pathologists scored TIL% of H&E slides of breast cancer from an external cohort (N = 199). TIL% was calculated referenced on the guideline of Immuno-Oncology Biomarker Working Group on Breast Cancer. For the cases of TIL score difference between each pathologist and AI model more than 15%, the pathologists were asked to revise TIL% in assistance with AI model which displays both stromal area and TIL. Finally, we compared the interobserver variation based on intraclass correlation coefficients (ICC) before and after AI assistance. Results The distribution of TIL score by 3 pathologists was 7% (5-20%), 15% (5-50%), and 20% (10-40%), respectively [median (25%-75% quantile)]. The ICC value of the initial TIL score evaluation was 0.716 (95% confidence interval, 0.560-0.811). Afterward, pathologists revised their initial scoring with assistance of AI model for the cases of difference more than 15% (n = 19, 72, and 73, respectively for each pathologist). After rescoring, number of slides with 15% or more difference of TIL% between raters significantly decreased from 109 slides (54.8%) to 75 slides (37.7%, p &lt; 0.001). The ICC value after re-scoring TIL% was 0.831 (95% confidence interval, 0.725-0.890). Conclusions There was a notable interobserver variation to score TIL% in breast cancer. Assistance with AI-powered TIL analyzer substantially improved the pathologist’s consensus and could be regarded as one of references for the final labeling of TIL%.

Citation Format: Soo Ick Cho, Wonkyung Jung, Sangjoon Choi, Seokhwi Kim, Sanghoon Song, Gahee Park, Minuk Ma, Seonwook Park, Sergio Pereira, Sangheon Ahn, Brian Jaehong Aum, Seunghwan Shin, Kyunghyun Paeng, Donggeun Yoo, Chan-Young Ock. Assistance with an artificial intelligence-powered tumor infiltrating lymphocytes (TIL) analyzer reduces interobserver variation in pathologic scoring of TIL in breast cancer [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P4-05-07.

921 Comparison of PI3K/AKT/mTOR pathway profiles amongst three immune phenotypes classified by artificial intelligence-powered H&E analyzer in non-small cell lung cancer

Background

The phosphatidylinositol 3-kinase (PI3K)/Akt/mechanistic target of rapamycin (mTOR) pathway plays a significant role in both tumorigenesis and progression of disease in non-small cell lung cancer (NSCLC).1 Increased activation of the pathway, whether in tumor or immune cells, results in an immunosuppressive tumor microenvironment.2 Therefore, we looked into how this pathway differs in three distinct NSCLC immune phenotypes.

Methods

Lunit SCOPE IO (Lunit, Seoul, Republic of Korea), a deep learning-based hematoxylin and eosin (H&E) image analytics tool, identifies lymphocytes and quantifies lymphocyte density within the cancer epithelium (CE-Lym), stroma (CS-Lym), and combined area (C-Lym). We applied Lunit-SCOPE IO to H&E-stained tissue images of 965 NSCLC samples from The Cancer Genome Atlas (TCGA). Tumors in the lowest tertile of C-Lym were labeled as immune-desert, and the remaining tumors were classified as inflamed and immune-excluded according to the median of the ratio of CE-Lym to CS-Lym.Utilizing RNA-sequencing data from TCGA, gene set enrichment analysis (GSEA) was conducted to analyze the differences in mTORC1 and PI3K/Akt/mTOR signaling between the subtypes.3 We obtained mutational data related to the PI3K/Akt/mTOR pathway from cBioPortal to compare the ratio of functional mutations between the immune phenotypes.4

Results

The mTORC1 signaling gene set was consistently enriched in immune-excluded, whether compared to inflamed (padj < 0.01, normalized enrichment score [NES]: 2.3) or immune-desert (padj < 0.01, NES: 1.6). However, PI3K/Akt/mTOR signaling gene set enrichment did not show statistically significant differences between the immune phenotypes.Within the three immune phenotypes, we analyzed three functional mutations: PIK3CA, PTEN, and Akt1 (figure 1). Of the total 112 samples showing the functional mutations of the PI3K/Akt/mTOR pathway, 53 were immune-excluded, 31 inflamed, and 28 immune-desert. The relation between mutation frequency and the immune subtypes was significant (X2 (2) = 11.1979, p < .01). The immune-excluded was more likely than the other subtypes to have functional PI3K/Akt/mTOR mutations.

Abstract 921 Figure 1

The landscape of functional mutation and immune phenotypes regarding PI3K/Akt/mTOR pathway

Conclusions

The three tissue phenomic subtypes showed different PI3K/Akt/mTOR pathway profiles, with immune-excluded having the most mutation samples and the greatest enhancement of mTORC1 signaling gene set. Likewise, tissue H&E based tumor microenvironment classification by Lunit SCOPE IO can be applied to other hallmark pathways and tumor types, and such further investigation of the tumor microenvironment can provide insights into novel therapeutic avenues.

References

Tan AC. Targeting the PI3K/Akt/mTOR pathway in non-small cell lung cancer (NSCLC). Thorac Cancer 2020;11(3):511–8.

O’Donnell JS, Massi D, Teng MWL, Mandala M. PI3K-AKT-mTOR inhibition in cancer immunotherapy, redux. Semin Cancer Biol 2018;48:91–103.

Liberzon A, Birger C, Thorvaldsdóttir H, Ghandi M, Mesirov JP, Tamayo P. The molecular signatures database hallmark gene set collection. Cell Systems 2015;1(6):417–25.

Cerami E, Gao J, Dogrusoz U, Gross BE, Sumer SO, Aksoy BA, et al. The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov 2012;2(5):401–4.

830 Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals immune-excluded phenotype related to APOBEC signature and clonal evolution of cancer

Background

Little is known about bridging clonal heterogeneity into the resistance of immune checkpoint inhibitors (ICI). Recent reports showed that excluded tumor-infiltrating lymphocytes (TIL) into stroma assessed by an artificial intelligence (AI)-powered spatial TIL analyzer, Lunit SCOPE IO, was related to loss-of-heterozygosity of HLA genes which would be one of crucial resistance pathways of ICI.1 In the current study, we hypothesized that Immune-excluded phenotype called by Lunit SCOPE IO would be related to clonal heterogeneity resulted from genome-wide accidents during early carcinogenesis which may cause an improper targeting of TIL for diverse clones with multiple genomic aberrations.

Methods

For spatial TIL analysis, we applied Lunit SCOPE IO1 which automatically detects TIL and segmentizes cancer area and stroma, then it classified Immune phenotype of 1 mm2-sized grid in H&E image. Inflamed score or Immune-excluded score were defined as the proportion of Inflamed phenotype, which is high intra-tumoral TIL density, or Immune-excluded phenotype, which is exclusively high TIL density only in stroma, within a whole-slide image, respectively. We evaluated the correlation of Immune phenotype with APOBEC mutational signature by single-base substitution (SBS) signature 2 and/or SBS13,2 whole-genome doubling, and subclonal genome fraction which reflects intra-tumoral heterogeneity,3 and clusters of T cell receptor (TCR) repertoire 4 derived from previous reports of The Cancer Genome Atlas (TCGA), consists of 7,467 tumor samples from 22 cancer types.

Abstract 830 Table 1

Correlation between immune phenotype and clonal evolution of cancer [* Median (95% confidence interval)]

Results

In the TCGA pan-carcinoma database, APOBEC mutational signature was significantly correlated with increased ratio of cancer stroma to cancer epithelium (median 0.866 vs 1.19, fold change +37.4%), and increased TIL density in cancer stroma (median 558 vs 764 / mm2, fold change +36.9%), but it was not correlated with intra-tumoral TIL density (median 63 vs 59 / mm2, fold change -6.3%). Interestingly, Immune-excluded score (IES) called by Lunit SCOPE IO was positively correlated with APOBEC mutational signature as well as expression levels of APOBEC1, APOBEC3A, and APOBEC3B, whole-genome doubling, and subclonal genome fraction, respectively, while Inflamed score (IS) or immune cytolytic activity (GZMA and PRF1 expressions) was negatively or not significantly correlated to those variables (table 1). TCR repertoire was expanded in the tumor samples with high IS (spearman rho = 0.279), but it was not increased in those with high IES (spearman rho = -0.0595).

Conclusions

There is a significant correlation between distinct TIL deposition in stroma, or Immune-excluded phenotype, with APOBEC-attributed clonal expansion of cancer, without proper expansion of TCR repertoire.

References

Ock CY, Park C, Paeng K, Yoo D, Kim S, Park S, Lee SH, Mok T, Bang YJ. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals distinct genomic profile of immune excluded phenotype in pan-carcinoma. Cancer Res 2021;81(Supp 13):1908.

Alexandrov LB, Kim J, Haradhvala NJ, Huang MN, Tian Ng AW, Wu Y, Boot A, Covington KR, Gordenin DA, Bergstrom EN, Islam SMA, Lopez-Bigas N, Klimczak LJ, McPherson JR, Morganella S, Sabarinathan R, Wheeler DA, Mustonen V, PCAWG Mutational Signatures Working Group, Getz G, Rozen SG, Stratton MR, PCAWG Consortium. The repertoire of mutational signatures in human cancer. Nature 2020;578(7793):94–101.

Taylor AM, Shih J, Ha G, Gao GF, Zhang X, Berger AC, Schumacher SE, Wang C, Hu H, Liu J, Lazar AJ, Cancer Genome Atlas Research Network, Cherniack AD, Beroukhim R, Meyerson M. Genomic and functional approaches to understanding cancer aneuploidy. Cancer Cell 2018;33(4):676–689.e3.

Zhang H, Liu L, Zhang J, Chen J, Ye J, Shukla S, Qiao J, Zhan X, Chen H, Wu CJ, Fu YX, Li B. Investigation of antigen-specific T-Cell receptor clusters in human cancers. Clin Cancer Res 2020;26(6):1359–1371.

823 Spatial analysis of tumor-infiltrating lymphocytes correlates with the response of metastatic colorectal cancer patients treated with vactosertib in combination with pembrolizumab

Background

Previously, we presented a promising anti-tumor efficacy (ORR: 16%, mOS: 15.8 months, RECIST) of the combination of vactosertib, a potent and selective TGF-β receptor I, and pembrolizumab (vac+pem) in patients with microsatellite stable metastatic colorectal cancer (MSS mCRC, MP-VAC-204 study). Recent reports showed immune-excluded TIL located in stroma would be closely related to TGF-β signature, which may harbor the primary resistance of pembrolizumab. In this study, we performed an exploratory biomarker analysis of TIL resided in either intra-tumoral or stromal area in pathology slides, and we hypothesized that spatial features of TIL would correlate with the response of vac+pem.

Methods

Pathology slides stained with H&E were obtained from 31 patients at baseline and 14 patients at cycle 2 in MSS mCRC patients in MP-VAC-204 study. For spatial TIL analysis, we applied an artificial intelligence -powered H&E analyzer, named Lunit SCOPE IO, which automatically detects TIL, tumor and stroma. It calculates the proportion of immune phenotype consists of inflamed, as high TIL density inside tumor area, or immune-excluded, as high TIL density in stroma in whole-slide images. Additionally, PD-L1 and CD8 were stained using multiplex immunohistochemistry to validate immune phenotype assessed by Lunit SCOPE IO.

Results

At baseline, the proportion of immune-excluded area (immune-excluded score, IES) was positively correlated with the density of CD8-positive cells in stroma area measured by mIHC (coefficient = 0.349), but it was not related to the density of PD-L1-positive cells (coefficient = -0.226). Area under receiver operating characteristics to predict the responder as partial response by RECIST v1.1 by IES and PD-L1 were 0.741 and 0.528. The overall response rate of vac+pem in the patients with high IES > 42.3% was 25% (4 out of 16), while no response was observed in those with low IES (0 out of 15). Overall survival (OS) of vac+pem was significantly prolonged in those with high IES > 42.3% compared to low IES (median OS: not reached versus 6.8 months, P = 0.0097), but it was not different according to PD-L1 level. After treatment of vac+pem, while IES was decreased regardless of treatment response, the proportion of inflamed area was increased in the responders (N=3) but decreased in the non-responders (N=11).

Conclusions

Immune-excluded score which reflects TGF-β-driven TIL exclusion into stroma is correlated with anti-tumor response of vac+pem in MSS mCRC. Further investigation on spatial TIL analysis as a potential biomarker should be warranted. (Clinical trial information: NCT03724851)

Abstract 617: Artificial intelligence-powered tissue analysis reveals distinct tumor-infiltrating lymphocyte profile as a potential biomarker of molecular subtypes in endometrial cancer

Introduction: Based on molecular classification of endometrial cancer (EC) of The Cancer Genome Atlas (TCGA) and Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE), EC has been classified into four novel prognostic groups: POLE-mutated (POLE-mt), mismatch repair-deficient (dMMR), copy number-high (p53abn), and copy number-low (no specific molecular profile, NSMP). We hypothesized that spatial distribution of tumor-infiltrating lymphocyte (TIL) using an artificial-intelligence (AI)-powered tissue analyzer, Lunit SCOPE, would be distinct according to the molecular classification.

Methods: We analyzed EC of TCGA database (N=224) and EC tissues retrospectively collected from Seoul National University Bundang Hospital (SNUBH, N=236). EC from SHUBH were molecularly classified in which MMR and p53 status were determined by immunohistochemistry (IHC) and POLE mutation by digital droplet polymerase chain reaction of six hotspot mutations in exon 9, 13 and 14 (P286R, S297F, V411L, V424I, L424V, and A456P). Lunit SCOPE analysis to detect lymphocyte, cancer epithelium (CE), and cancer-associated stroma (CS) was performed using x40 scanned images of TCGA and SNUBH. Cox proportional hazard model were used for survival analysis.

Results: Composition of molecular classification were comparable in both cohorts: 7.6%, 28.6%, 24.6%, and 39.3% for POLE-mt, dMMR, CN-high, and CN-low in TCGA cohort, and 8.9%, 19.5%, 17.4%, and 54.2% for POLE-mt, dMMR, p53abn, and NSMP in SNUBH cohort, respectively. CN-high subtype in TCGA and p53abn in SNUBH cohort significantly correlated with poor prognosis (TCGA: adjusted hazard ratio 2.54, p value 0.0383; SNUBH: adjusted hazard ratio 7.47, p value 2.38 x 10-4). TIL density calculated by total number of lymphocytes in CE and CS area was significantly increased in POLE and MSI groups of TCGA-cohort (p value = 2.24 x 10-4) and SNUBH-cohort (p value = 1.74 x 10-6). Moreover, uneven TIL enrichment in CS was observed in CN-high or TP53-expressor compared to CN-low or TP53-wildtype (ratio of TIL in CS/TIL in CE, TCGA: 6.96 versus 5.33; SNUBH: 11.4 versus 8.83). These findings suggested that the CN-high/p53abn and POLE/dMMR subtypes might be associated with immune-excluded and immune-inflamed tumor microenvironment (TME) phenotype, respectively.

Conclusion:The distribution of TIL in EC differs according to the molecular subtypes. Our data suggest the possibility of predicting subtypes through TIL analysis, and provide insight into treatment through TME modulation.

Citation Format: Hyojin Kim, Eun Sun Kim, Song Kook Lee, Jeong Hoon Lee, Kyunghyun Paeng, Chan-Young Ock, Dong Hoon Suh, Kidong Kim, Jae Hong No, Yong-Beom Kim. Artificial intelligence-powered tissue analysis reveals distinct tumor-infiltrating lymphocyte profile as a potential biomarker of molecular subtypes in endometrial cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 617.

Abstract 1908: Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals distinct genomic profile of immune excluded phenotype in pan-carcinoma

Introduction: Immune-excluded phenotype, defined by the existence of tumor-infiltrating lymphocytes (TIL) exclusively confined to cancer-associated stroma (CS) without protruding to tumor nest, has been suggested to be an intrinsic resistance mechanism of immune checkpoint inhibitor. However, little is known about the genomic landscape of immune-excluded phenotype across cancer types. In the current study, we analyzed genomic correlates of immune excluded phenotype in pan-carcinoma, using Lunit SCOPE IO, an artificial intelligence (AI)-powered whole slide image (WSI) software analyzer.

Methods: Lunit SCOPE IO has been trained for &gt; 100,000 WSI across &gt; 30 cancer types to detect lymphocyte, cancer epithelium (CE), and CS with &gt; 95% accuracy. WSI was divided into 1 mm2 sized patch and lymphocyte density in CE or its density in CS were calculated to classify immune phenotype as follows: lymphocyte infiltration into CE was considered as inflamed, lymphocyte exclusively enriched in CS as immune-excluded, and sparse lymphocyte infiltration as immune-desert. We analyzed 7,128 The Cancer Genome Atlas (TCGA) pan-carcinoma samples across 20 cancer types excluding those of mesenchymal origin.

Results: Lunit SCOPE IO classifies WSI of TCGA pan-carcinoma into three immune phenotypes: inflamed (17.9%), immune-excluded (27.7%), immune-desert (34.2%) and remaining 20.2% of mixed phenotype. Immune-excluded phenotype is highly enriched in lung squamous cell carcinoma (61.5%), lung adenocarcinoma (53.6%), colorectal cancer (51.7%), and pancreatic cancer (42.2%) whereas inflamed phenotype is less dominant in colorectal cancer (3.4%) and pancreatic cancer (3.3%). Tumor mutational burden (TMB) is increased in both inflamed (mean 7.30/Mb) and immune-excluded (7.18/Mb) compared to that in immune-desert (3.28/Mb), however, genomic instability assessed by fraction altered loss-of-heterozygosity (LOH) is only increased in immune-excluded (mean 0.208) compared to that in inflamed (0.143) and immune-desert (0.148, P &lt; 10-16). Consistent with this result, tumors with high TMB (&gt; 10/Mb) and LOH of HLA genes (N = 124) is predominantly enriched by immune-excluded (45.2%), compared to inflamed (23.4%) or immune-desert (8.9%). Interestingly, a majority of oncogenic drivers such as mutations of TP53, KRAS, and KEAP1 are significantly enriched in immune-excluded (fold change &gt; 1.5, P &lt; 10-10), and gene sets associated with epithelial-mesenchymal transition, apical junction, and TGF-beta signaling are enriched in immune-excluded.

Conclusion: AI-powered spatial analysis of WSI can classify immune phenotype in pan-carcinoma and it reveals that plausible immune intrinsic resistance pathways including genomic instability, LOH of HLA genes, and alteration of major oncogenic drivers are highly enriched in immune-excluded phenotype.

Citation Format: Chan-Young Ock, Changhee Park, Kyunghyun Paeng, Donggeun Yoo, Seokhwi Kim, Sehhoon Park, Se-Hoon Lee, Tony S. Mok, Yung-Jue Bang. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes reveals distinct genomic profile of immune excluded phenotype in pan-carcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 1908.

Interim analysis of first-in-human phase 1 study to assess safety and efficacy of YBL-006, an anti-PD-1 antibody in advanced solid tumor with exploratory biomarker analysis of tumor mutational burden and artificial intelligence (AI)-powered spatial analysis of tumor-infiltrating lymphocytes

e14552

Background: YBL-006 is an anti-programmed death-1 (PD-1) antibody with a higher affinity compared to that of other PD-1 antibodies, which showed a favorable safety profile in animal models. We designed the first-in-human phase I trial of YBL-006 to assess its safety and efficacy with exploratory biomarker analysis in patients with advanced solid tumors refractory to standard of treatment. Methods: A modified “3+3” design, with the first patient dosed at 0.5 mpk, was followed by conventional dose escalation of 2, 5, and 10 mpk IV. Pharmacokinetics (PK) and pharmacodynamics, including PD-1 receptor occupancy (RO) and serum levels of interferon-gamma (IFN-γ), were assessed. Adverse events (AEs) were graded using the CTCAE v4.03. Tumor response was assessed using the RECIST v1.1 every 8 weeks. For exploratory analysis, tumor mutational burden (TMB) and AI-powered spatial analysis of tumor-infiltrating lymphocyte (TIL) of tumor tissues collected before YBL-006 treatment were performed. The cut-off date for analysis was February 12, 2021. Results: A total of 8 patients enrolled in the 0.5, 2, and 5 mpk cohorts received at least one dose of YBL-006 and median exposure was 15 weeks (ranged 4-26). No dose limiting toxicity occurred and the maximum tolerated dose was not reached until progressing to the 5 mpk. The common treatment-related AEs were G1 fatigue (25%), and G1 hypothyroidism (12.5%). We also observed 1 case of G2 cytokine release syndrome during cycle 1 in 2 mpk which was managed with supportive care alone. No treatment-related deaths have occurred to date. YBL-006 showed a linear PK prolife and both PD-1 RO and serum IFN-γ increased by > 2 times 8 h after the first dose. Tumor evaluation data were available for 7 patients, which showed 1 confirmed complete response (CR, penile squamous cell carcinoma, 2 mpk) and 1 confirmed partial response (PR, anal squamous cell carcinoma, 2 mpk) with durable responses lasting more than 19+ and 10+ weeks respectively, 2 stable disease (SD) and 3 progressive disease (PD). Four tumor samples were available for biomarker analysis. TMBs of patients with CR (8.3/Mb) or PR (9.3/Mb) were higher than those in 2 patients with PD (5.5 and 1.7/Mb). AI-powered spatial analysis of TIL showed that intratumoral TIL density was increased in patients who achieved CR and PR (66.1% and 95.8%, respectively) compared to those in patients who exhibited PD (25.1% and 16.5%, respectively). Conclusions: Interim analysis of phase I study showed that YBL-006 is well tolerated and preliminary biomarker analysis showed that the TMB, and intratumoral TIL infiltration are potentially related to the response to YBL-006. Clinical trial information: NCT04450901.

Distinct subset of immune cells assessed by multiplex immunohistochemistry correlates with immune phenotype classified by an artificial intelligence-powered tissue analyzer in advanced non-small cell lung cancer

e21012

Background: Pathologic classification of immune phenotype is challenging since there is no consensus on how to assess spatial relations of tumor-infiltrating lymphocyte (TIL) on cancer epithelium (CE) and cancer stroma (CS) in whole-slide images (WSI). We previously suggested that the artificial intelligence (AI)-powered tissue analyzer, Lunit SCOPE IO, can classify immune phenotype, and that its predictions are correlated with the clinical outcome of immune checkpoint inhibitor (ICI) in non-small cell lung cancer (NSCLC). In this study, we designed a pathologic validation of immune phenotype using multiplex immunohistochemistry. Methods: Lunit SCOPE IO was developed based on a 2.8 x 109 micrometer2 area of CE or CS, and 5.9 x 106 TILs from 3,166 H&E Whole-Slide Image (WSI) of multiple cancer types, annotated by board-certified pathologists. H&E WSIs were divided into 1 mm2-sized tiles, where we classified immune phenotype (IP) based on TIL density on CE and TIL density on CS. Representative IP was determined based on the overall proportion of tile-level IPs in each WSI. Multiplex immunohistochemistry (mIHC) staining with CD3, CD8, CD20, CD68, FOXP3, CK, and DAPI was performed in NSCLC tumor tissues (n = 99) treated with immune checkpoint inhibitors (ICI) at the Samsung Medical Center. A normalized number of cells expressing each marker was calculated by dividing the total number of marker-positive cells by the number of DAPI-positive cells in each WSI. Results: The proportions of inflamed IP, immune-excluded IP, and immune desert IP in the analysis set were 46.5%, 29.3%, and 24.2%. respectively. Median progression-free survival of ICI was 6.4 m in inflamed IP, 1.9 m in immune-excluded IP, and 1.6 m in immune-desert IP (hazard ratio of inflamed versus others: 0.43, confidence interval 0.27-0.68, P = 0.000188). Multiplex IHC results showed that the normalized CD3-positive cells and CD8-positive cells, which play a role of anti-tumor activity, were highly enriched in inflamed IP compared to those in other IPs (CD3: fold change [FC] 1.57, P = 0.0182; CD8: FC 1.24, P = 0.0697), whereas FOXP3-positive cells, linked to the immunosuppressive activity, were enriched in immune-excluded IP (FC 1.26, P = 0.0656). We also noted that CD68-positive cells were significantly enriched in immune-desert IP (FC 1.76, P = 0.00467). Conclusions: The immune cell subset in WSI is distinct according to the immune phenotype, as CD3- or CD8-positive cells are enriched in inflamed IP rather than immune-excluded IP as classified by AI-powered TIL analysis of H&E image.

Pathologic validation of artificial intelligence-powered prediction of combined positive score of PD-L1 immunohistochemistry in urothelial carcinoma

e16518

Background: Programmed death ligand 1 (PD-L1) expression is a reliable biomarker of immune-checkpoint inhibitors (ICI) in multiple cancer types including urothelial carcinoma (UC). A 22C3 pharmDx immunohistochemistry was particularly determined by using the combined positive score (CPS) in UC. A challenging issue regarding the manual scoring of CPS by a pathologist is in determining the representative area to read. This requires substantial time and effort and may lead to inter-observer variation. We developed an artificial intelligence (AI)-powered CPS analyzer, to assess CPS in whole-slide images (WSI) and validated its performance by comparing against a consensus of pathologists’ readings. Methods: An AI-powered CPS analyzer, Lunit SCOPE PD-L1, has been trained and validated based on a total of 3,326,402 tumor cells, lymphocytes, and macrophages annotated by board-certified pathologists for PD-L1 positivity in 1200 WSI stained by 22C3. After excluding the in-house control tissue regions, the WSIs were divided into patches, from which a deep learning-based model was trained to detects the location and PD-L1 positivity of tumor cells, lymphocytes, and macrophages, respectively. Finally, the patch-level cell predictions were aggregated for CPS estimation. The performance of the model was validated on an external validation UC cohort consisting of two institutions: Boramae Medical Center (BMC, n = 93) and Seoul National University Bundang Hospital (SNUBH, n = 100). Three uropathologists independently annotated the CPS of the external validation cohorts, and a consensus of CPS was determined by determination of their mean values. Results: The AI-model predicts CPS accurately in an internal validation cohort as the area under the curves (AUC) values to predict PD-L1-positive tumor cell, PD-L1-positive lymphocytes or macrophages, PD-L1-negative tumor cell, and PD-L1-negative lymphocytes or macrophages were 0.929, 0.855, 0.885, and 0.872, respectively. There was a significant positive correlation between CPS by AI-model and consensus CPS by 3 pathologists in the external validation cohort (Spearman coefficient = 0.914, P < 0.001). Concordance of AI-model and pathologists' consensus to call CPS ≥ 10 was 88.1%, which was similar to that of either 2 of 3 pathologists (84.5%, 86.5%, and 90.7%). The concordance rate was not significantly different according to data source (BMC: 88.2% versus SNUBH: 88.0%, P = 1.00), but was significantly different according to type of surgery [surgical resection (cystectomy, nephrectomy, and ureterectomy): 92.3% versus transurethral resection: 81.3%, P = 0.0244]. Conclusions: Lunit SCOPE PD-L1, AI-powered CPS analyzer, can detect PD-L1 expression in tumor cells, lymphocytes or macrophages highly accurately compared to uropathologists.

Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes predicts survival after immune checkpoint inhibitor therapy across multiple cancer types

2607

Background: Tumor infiltrating lymphocytes (TIL) are a potential tumor-agnostic biomarker for immune checkpoint inhibitor (ICI) therapy. We previously reported the clinical application of an artificial intelligence-powered spatial TIL analyzer, Lunit SCOPE IO, for predicting ICI treatment outcomes in advanced non-small cell lung cancer (NSCLC). Here, we expand the clinical application of Lunit SCOPE IO as a tumor-agnostic ICI biomarker across multiple cancer types. Methods: Lunit SCOPE IO was trained and validated with a 2.8 x 109 micrometer2 area and 5.9 x 106 TILs from 3,166 H&E Whole-Slide Images (WSI) of multiple cancer types, annotated by 52 board-certified pathologists. The Inflamed Score (IS) was defined as the proportion of all tumor-containing 1 mm2-size tiles within a WSI classified as being of the inflamed immune phenotype (high TIL density within cancer epithelium). We first evaluated the correlation between the IS and TMB, MSI-H, and immune cytolytic activity ( GZMA and PRF1) across 22 cancer types from The Cancer Genome Atlas (TCGA, n = 7,467). Subsequently, the correlation between the IS and overall survival after ICI treatment was evaluated in a real-world dataset of patients with 9 different tumor types (n = 1,013), retrospectively collected from Stanford University Medical Center, Chonnam National University Hospital, Samsung Medical Center, and Seoul National University Bundang Hospital. Results: Lunit SCOPE IO accurately detected CE, CS, and TILs with an area under the receiver-operating-characteristic curve of 0.970, 0.949, and 0.925, respectively. In the TCGA pan-cancer cohort, Lunit SCOPE IO’s IS correlated significantly with immune cytolytic activity (Spearman rho = 0.504, p < 0.001), TMB-high (≥ 10 mutations/Mb, fold change 1.39, p < 0.001) and MSI-H (fold change 1.45, p < 0.001). The IS-positive proportions of microsatellite-stable (MSS) and TMB-low cases were 42.5% and 17.1%, using the thresholds of IS ≥ 20% and ≥ 50% as presumptive clinical cutoffs. In the real-world ICI clinical dataset (n = 1,013), an IS ≥ 20% correlated significantly with favorable overall survival after ICI treatment (cancer type-adjusted hazard ratio [HR] 0.70, 95% confidence interval [CI] 0.59-0.83, p < 0.0001). Furthermore, this association remained significant after the exclusion of NSCLC patients (n = 519) (adjusted HR 0.68, 95% CI 0.53-0.86, p = 0.0016) indicating that the effect was not driven solely by one major tumor type. Conclusions: The Inflamed Score (IS), as evaluated by Lunit SCOPE IO, correlates with favorable overall survival after ICI treatment across multiple tumor types. AI-powered spatial TIL analysis of the tumor microenvironment may be able to detect a significant proportion of ICI responders, and offers promise as a new companion diagnostic, particularly in patients with MSS/TMB-low tumors.

Clinical performance of artificial intelligence-powered annotation of tumor cell PD-L1 expression for treatment of immune-checkpoint inhibitor (ICI) in advanced non-small cell lung cancer (NSCLC)

9026

Background: Programmed death ligand 1 (PD-L1) expression is the standard biomarker for first line ICI in advanced NSCLC. However, manual evaluation of tumor proportion score (TPS) by pathologists has practical limitations including intra/inter-observer bias, variation in subjectivity on area of interest and intensive labor. We developed an artificial intelligence (AI)-powered TPS analyzer, namely Lunit SCOPE PD-L1, for objective annotation of tumor cell PD-L1 expression for prediction of ICI response in advanced NSCLC. Methods: Lunit SCOPE PD-L1 was developed by a total of 393,565 tumor cells annotated by board-certified pathologists for PD-L1 expression in 802 whole-slide images (WSI) stained by 22C3 pharmDx immunohistochemistry. A After excluding the in-house control tissue regions, the WSI were divided into patches, from which a deep learning-based model detected the location and PD-L1 positivity of tumor cells. The patch-level cell predictions were aggregated for TPS estimation. Clinical performance of the model was validated in an external cohort of 430 NSCLC tumor slides from patients treated with ≥ ICI at Seoul National University Bundang Hospital and Samsung Medical Center. Independent control TPS annotation of this external validation cohort was performed by three pathologists, and their consensus TPS was calculated by mean value of such. Results: AI-model (Lunit SCOPE PD-L1) predicts PD-L1-positive tumor cell with the area under the curves of 0.889 and PD-L1-negative tumor cells with that of 0.809 at cell-level analysis. At WSI-level, significant positive correlation was observed between TPS by AI model and control TPS by pathologists (Spearman coefficient = 0.9247, P < 0.001). Concordance rate between AI-model and control TPS by pathologists according to expression level of PD-L1 ≥ 50%, 1-49%, and < 1% status was 85.7%, 89.3%, and 52.4%, respectively. Median progression-free survival (mPFS) according to TPS by AI model ≥ 1% vs. < 1% were 2.8 vs. 1.7 months (hazard ratio, HR, 0.52, 95% confidence interval, CI, 0.38-0.71, P < 0.001). In contrast, mPFS according to control TPS was 2.8 vs. 2.1 months (HR 0.70, 95% CI 0.55-0.91, P < 0.001). Forty out of 84 patients (47.6%) annotated as control TPS < 1% by pathologists were considered as TPS ≥ 1% by AI-model and mPFS of this subgroup was 2.7 months. Conclusions: PD-L1 expression by AI-model correlates with PD-L1 expression by pathologists. Clinical performance of AI-model in WSI-level is comparable with assessment by pathologists. The AI-model can accurately predict tumor response and progression-free survival of ICI in advanced NSCLC.

Artificial intelligence-powered spatial analysis of tumor infiltrating lymphocytes (TIL) to reflect target gene expressions of novel immuno-oncology agents

e14534

Background: Novel immuno-oncology (IO) agents are promising but showing their efficacy in early phase clinical trials has been challenging due to limited enrichment strategies using practical biomarker platforms. We hypothesize that an artificial intelligence (AI)-powered spatial analysis of TIL using practically feasible H&E slides, can reflect a specific target gene expression derived from RNA sequencing. This enhances its potential application in early development of novel IO agents. Methods: An AI-powered spatial TIL analyzer, namely Lunit SCOPE IO, was developed with data from 2.8 x 109 micrometer2 H&E-stained tissue regions and 5.9 x 106 TILs from 3,166 whole slide images of multiple cancer types, annotated by board-certified pathologists. Inflamed Score and Immune-Excluded Score was defined as the proportion of all tumor-containing 1 mm2-size tiles within a WSI classified as being of inflamed immune phenotype (high TIL density within cancer epithelium) and immune-excluded phenotype (low TIL density within cancer epithelium, but high TIL density within stroma), respectively. We used RNA sequencing data and H&E images from The Cancer Genome Atlas database, excluding those of mesenchymal origin (n = 7,467). Spearman's rank correlation between each gene expression and IS or IES, respectively, was calculated. Correlation coefficient > 0.2 and false discovery rate (FDR) < 1% was considered as a significant correlation. Results: In a total of 20,304 genes, 871 (4.3%) and 1,155 (5.7%) genes were significantly correlated with Inflamed Score (IS) and Immune-Excluded Score (IES), respectively. The IS was highly related to genes reflecting immune cytolytic activity and targets of approved immune checkpoint inhibitors (Table). Interestingly, it was also significantly correlated with target genes of novel IO such as TIGIT, LAG3, TIM3, IDO, Adenosine receptor A2A, OX40, ICOS, M-CSF, IL2, IL7, and IL12. Moreover, the IES was exclusively correlated with the target genes of CEACAM, TGFB, and IL1. Conclusions: Expression levels of novel I-O target genes are correlated with three scores derived from AI-powered TIL analysis using H&E slides, which can be easily applied to clinical research.[Table: see text]

Accuracy and efficiency of an artificial intelligence tool when counting breast mitoses

BACKGROUND

The mitotic count in breast carcinoma is an important prognostic marker. Unfortunately substantial inter- and intra-laboratory variation exists when pathologists manually count mitotic figures. Artificial intelligence (AI) coupled with whole slide imaging offers a potential solution to this problem. The aim of this study was to accordingly critique an AI tool developed to quantify mitotic figures in whole slide images of invasive breast ductal carcinoma.

METHODS

A representative H&E slide from 320 breast invasive ductal carcinoma cases was scanned at 40x magnification. Ten expert pathologists from two academic medical centers labeled mitotic figures in whole slide images to train and validate an AI algorithm to detect and count mitoses. Thereafter, 24 readers of varying expertise were asked to count mitotic figures with and without AI support in 140 high-power fields derived from a separate dataset. Their accuracy and efficiency of performing these tasks were calculated and statistical comparisons performed.

RESULTS

For each experience level the accuracy, precision and sensitivity of counting mitoses by users improved with AI support. There were 21 readers (87.5%) that identified more mitoses using AI support and 13 reviewers (54.2%) that decreased the quantity of falsely flagged mitoses with AI. More time was spent on this task for most participants when not provided with AI support. AI assistance resulted in an overall time savings of 27.8%.

CONCLUSIONS

This study demonstrates that pathology end-users were more accurate and efficient at quantifying mitotic figures in digital images of invasive breast carcinoma with the aid of AI. Higher inter-pathologist agreement with AI assistance suggests that such algorithms can also help standardize practice. Not surprisingly, there is much enthusiasm in pathology regarding the prospect of using AI in routine practice to perform mundane tasks such as counting mitoses.

Deep learning-based immune phenotype analysis reveals distinct resistance pattern of immune checkpoint inhibitor in non-small cell lung cancer

3119

Background: Resistance pattern and biological mechanism of immune checkpoint inhibitor (ICI) has been poorly understood. Sine suggested resistance mechanisms would be either innate resistance caused by lack of immune recruitment or acquired immune evasion after durable response of ICI treatment, we hypothesized that resistance pattern of tumor microenvironment would be distinct according to duration of ICI response in non-small cell lung carcinoma (NSCLC). In the current study, we applied deep-learning-based classification of three immune phenotypes (3IP): inflamed, excluded, and desert, to objectively assess the immunologic status of tumor microenvironment. Methods: Deep-learning algorithm of H&E Whole-Slide Images (WSI), called Lunit-SCOPE, was trained with 1,824 H&E WSI of NSCLC from Samsung Medical Center (SMC). WSI was divided into patches and each patch (~10 high-power fields) was classified as inflamed, excluded and desert, based on both quantity and localization of immune cells. Among NSCLC patients treated with ICI in SMC, 87 paired treatment-naïve (Pre, patch N = 15,415) and post-progression (Post, patch N = 18,197) tumor tissues were analyzed for Lunit-SCOPE. Results: In 87-paired samples, proportions of excluded and desert phenotypes were increased in post-progression tumor tissues (excluded; Pre 26.8% versus Post 32.5%, desert; Pre 19.5% versus Post 25.3%). Focused on 29 patients classified as inflamed in treatment-naïve, proportion of immune phenotypes of post-progression were clearly different according to duration of response, divided by median progression-free survival (PFS) of 3.7 m. Patients with rapid progression without ICI response (PFS < 3.7 m) turned into desert type (46.2%), whereas durable responder (PFS ≥ 3.7 m) either still remained on inflamed phenotype (42.9%) or turned into excluded phenotype (21.4%). Patients who remained on inflamed phenotype had favorable overall survival after progression on ICI, compared to turned into desert type (median survival not reached versus 6.6 m, P= 0.0296). Conclusions: Resistance patterns of ICI are distinct according to duration of response in patients with inflamed phenotype. Rapid progressor turns off immune into desert phenotype whereas most durable responder keeps immune recruitment into tumor microenvironment, which needs tailored strategy to overcome ICI resistance.

Deep-learning analysis of H&E images to define three immune phenotypes to reveal loss-of-target in excluded immune cells as a novel resistance mechanism of immune checkpoint inhibitor in non-small cell lung cancer

3120

Background: Discovery of predictive biomarker to enrich the responder of immune checkpoint inhibitor (ICI) in PD-L1-low ( < 50%) non-small cell lung cancer (NSCLC) is still challenging. Recent study showed that loss of heterozygosity (LOH) of HLA led to immune evasion. In the current study, we hypothesized that 3 immune phenotype (3IP): inflamed, excluded and desert would be reliably classified by deep-learning algorithm of H&E image, called Lunit-SCOPE, which would dictate the responder in PD-L1-low NSCLC patients and discover a unique resistance pathway in excluded phenotype. Methods: Lunit-SCOPE was trained with 1,824 H&E Whole-Slide Image (WSI) of NSCLC from Samsung Medical Center (SMC). WSI was divided into patches (~10 high-power fields) which was classified for 3IP, based on both quantity and localization of immune cells. The 3IP was trained and optimized by considering clinical outcome of 119 NSCLC patients with PD-(L)1 inhibitor (training cohort, patches = 25,897), and validated in 62 patients enrolled in LC-biomarker study (NCT03578185, validation cohort, patches = 8,929). Tumor Proportion Score (TPS) of PD-L1 22C3 immunohistochemistry was assessed by pathologists. Tumor Mutational Burden (TMB) was calculated as number of nonsynonymous alterations throughout whole-exome and HLA LOH was called by LOHHLA algorithm. Results: Interactive analysis to classify 3IP in training cohort showed that 8,726 (33.7%), 10,965 (42.3%), and 6,206 (24.0%) patches were classified as inflamed, excluded, and desert, respectively. In validation cohort, median progression-free survival (mPFS) of inflamed phenotype was 10.1 m, significantly prolonged compared to either excluded phenotype (3.0 m, P= 0.0053) or desert phenotype (1.4 m, P= 0.0011). Inflamed phenotype independently dictated favorable ICI outcome in PD-L1-low (TPS < 50%, mPFS of inflamed: 14.3 m vs excluded/desert: 1.4 m, P= 0.0233) as well as in PD-L1-high (TPS≥50%, 10.1 m vs 4.2 m, P= 0.0361), respectively. Excluded phenotype had higher TMB compared to inflamed phenotype had (median 177 vs 107), and HLA LOH was also enriched in excluded phenotype (31.0%) compared to inflamed (17.6%) and desert (16.7%) phenotypes. Conclusions: Lunit-SCOPE based 3IP classification can predict ICI outcome especially in PD-L1-low ( < 50%) patients. Excluded phenotype showed poor ICI outcome even with high TMB, partially explained by HLA LOH resulting in loss-of-target, as a novel resistance mechanism of ICI.

Deep-learning analysis of CT imaging biomarker for PD-L1 expression to predict heterogeneous response to immune checkpoint inhibitors in non-small cell lung carcinoma

e21529

Background: Inter-tumoral genomic heterogeneity cause various immune status in tumor microenvironment, which may lead to indiscriminate response to immune checkpoint inhibitor (ICI) in patients with multiple lesions. Therefore, PD-L1 status from practically approachable, single lesion would not be always representative for immune status of whole metastatic lesions of a patient. To solve inter-tumoral heterogeneity issue before biopsy in clinic, we developed a deep-learning based CT biomarker for predicting PD-L1 status, then explored if the algorithm would also predict ICI response of multiple lesions in non-small cell lung carcinoma (NSCLC). Methods: Deep learning-based image analyzer was trained with CT images of NSCLC in Samsung Medical Center (SMC) (N = 104). Taking 3D patch of a lesion located by radiologists, 3D convolutional neural network was trained to predict PD-L1 (22C3 immunohistochemistry) tumor proportion score of each lung or lymph node lesion. The prediction model was validated using publicly available dataset (NSCLC Radiogenomics, N = 115). Finally, we applied the model to baseline CT who had multiple lesions (≥ 2) and also received ICI (SMC validation, N = 170). Tumor response was assessed based on RECIST 1.1, and discordant response was defined by best response of each lesion outside -10% ~ +10%. Results: Predicted PD-L1 score was positively correlated with real PD-L1 expression in NSCLC Radiogenomics (Pearson = 0.198, P= 0.0339). In SMC validation cohort, predicted PD-L1 score of each lesion was negatively correlated with ICI response of corresponding lesion (Pearson = -0.0941, P= 0.0325). Interestingly, 35 out of 170 (20.6%) patients showing discordant ICI response among lesions had worse progression free survival (PFS) (median PFS: 2.1m, 1.6m, 2.5m, and 18.7m in discordant response, concordant progress, stable, and regress, respectively, P< 0.001). Patients with discordant response had significantly wide-ranged predicted PD-L1 score compared with others (median range: 0.273 versus 0.185, P= 0.0079). Moreover, patient-level median of predicted PD-L1 scores of lesions was significantly associated with prolonged PFS (hazard ratio 0.69, P= 0.0401) and overall survival (hazard ratio 0.65, P= 0.0431). Conclusions: Deep-learning based imaging biomarker accurately predicts PD-L1 status of each metastasis, as well as independent ICI response reflecting inter-tumoral heterogeneity. This algorithm would guide which lesion would be representative in clinic.

Abstract 3144: Deep learning-based predictive biomarker for adjuvant chemotherapy in early-stage hormone receptor-positive breast cancer

Introduction: Predictive value of adjuvant chemotherapy for patients with early-stage hormone receptor-positive breast cancer has been suggested by 21-gene expression assay, although its cost-effectiveness has not been well-defined. We have developed the deep learning-based H&E image analyzer named Lunit SCOPE, identifying and quantifying various histologic components from H&E-stained whole slide images We hypothesized that cell proportions analyzed by Lunit SCOPE would be a potential prognostic and predictive biomarker of adjuvant chemotherapy in early-stage hormone receptor-positive breast cancer.

Method: We have collected clinical data and H&E slides from de-identified 2,915 early breast cancer patients in Samsung Medical Center, retrospectively. The 898 patients with hormone receptor-positive, T1b ~ T3 and N0 ~ N1mi have been selected to analyze the predictive value of adjuvant chemotherapy. Deep learning-based H&E image analyzer, Lunit SCOPE, has been trained by 1,191 H&E-stained whole slide images from another breast cancer patient cohort. In the whole slide image, biological and histological components such as cancer epithelium, cancer stroma, normal, fat, necrosis, lymphocyte, fibroblast, and other cells, have been annotated by over 10 pathologists. The outputs of Lunit SCOPE are the ratio of cancer epithelium, cancer stroma, normal, necrosis and fat in a whole slide image and intratumoral tumor infiltrating lymphocyte (TIL) and stromal TIL density. The recurrence score (RS) based on the output of Lunit SCOPE has been determined by using multivariate cox regression analysis for disease-free survival (DFS) in the patients without adjuvant chemotherapy.

Result: Recurrence score (RS) was proportional to the cancer stroma ratio and stromal TIL density, but inversely proportional to intratumoral TIL density. When the RS cutoff was 0.913, 21.3% (191 out of 898) of patients were classified as high risk group (RS &gt; cutoff). Among those without adjuvant chemotherapy, high risk group presented poor DFS (hazard ratio [HR] 4.23, 95% confidence interval [CI] 1.87-9.59, P = 1.73 x 10-4) and overall survival (OS, HR 4.95, 95% CI 1.39-17.6, P = 6.07 x 10-3) than the low risk group. Adjuvant chemotherapy did not prolong OS in patients with low risk group (HR 1.08, 95% CI 0.38-3.12, P = 0.885). However, interestingly, in those with high risk by Lunit SCOPE, adjuvant chemotherapy prolonged DFS (HR 0.35, 95% CI 0.15-0.86, P = 0.0161) and OS (HR 0.22, 95% CI 0.05-0.95, P = 0.0254), reflecting RS by Lunit SCOPE would be a significant predictive biomarker of adjuvant chemotherapy.

Conclusion: Deep learning-based H&E image analyzer, Lunit SCOPE, was possible to analyze the prognosis of breast cancer. Especially, only high risk patients of RS by Lunit SCOPE had survival benefit from adjuvant chemotherapy, which needs to be validated in clinical trials.

Citation Format: Soo Youn Cho, Eun Yoon Cho, Kyunghyun Paeng, Geunyoung Jung, Sarah Lee, Sang Yong Song. Deep learning-based predictive biomarker for adjuvant chemotherapy in early-stage hormone receptor-positive breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 3144.

Deep learning-based predictive biomarker for immune checkpoint inhibitor response in metastatic non-small cell lung cancer

9094

Background: In the era of immunotherapy, immune checkpoint inhibitor (ICI) has changed the treatment paradigm in metastatic non-small cell lung cancer (NSCLC). Along with clinical trials, there is an ongoing investigation to discover the predictive biomarker of ICI which so far has unsatisfactory reliability. As an effort to enhance the predictive value of ICI treatment, we applied deep learning and developed artificial intelligent (AI) score (range from 0 to 1) to analyze the specific context of immune-tumor microenvironment (TME) extracted by scanned images from H&E slides. Methods: As a ground work, deep learning-based H&E image analyzer, Lunit SCOPE, has been trained with H&E images (n = 1824) from ICI naive NSCLC samples. For the calculation of AI score, training was conducted using responder/non-responder labeled ICI treated samples from the exploratory cohort. The ICI responder was defined as the patient with a best overall response of partial or complete response and stable disease for more than 6 months. The positivity of PD-L1 immunohistochemistry (IHC) was assessed manually by pathologists. Results: The exploratory cohort is composed of NSCLC patients treated with ICI (n = 189) in Samsung Medical Center, and response to ICI was observed in 72 (38.1%) patients. Median follow-up duration was 6.8 months (6.6~8.2). Samples with PD-L1 IHC positive, defined by ≥ 1%, was observed in 138 (73.0%) patients. AI score was significant higher in the responder group (median: 0.391 vs 0.205, P = 6.14e-5), and the patients with AI score above the cut-off (0.337) showed a better response to ICI (odds ratio [OR] 3.47 P = 7.34e-5) which is higher than patients with PD-L1 ≥ 1% (OR 1.92, P = 0.069). High AI score group (n = 83) showed significantly favorable PFS compared to low AI score group (n = 106, median PFS: 5.1m vs 1.9m, hazard ratio [HR] 0.51, P = 9.6e-5) and this outcome was independent with PD-L1 status (P = 6.0e-5). In subgroup analysis, PFS of PD-L1 high / AI score high group (n = 63) had longer median PFS (6.7m) compared to both PD-L1 high / AI score low group (n = 70, 4.0m, P = 0.001) and PD-L1 low/AI score low group (n = 35, 1.9m, P = 4.0e-6). Tumor infiltrating lymphocyte (TIL) density of cancer epithelium was significantly correlated with AI score (Pearson’s r = 0.310, P = 1.43e-5), which suggests that AI score may partly reflect TME represented by TIL. Conclusions: The AI score by machine-learned information, extracted from H&E images without additional IHC stain, could predict responsiveness and PFS of ICI treatment independent of PD-L1 IHC positivity.

From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The CAMELYON17 Challenge

Automated detection of cancer metastases in lymph nodes has the potential to improve the assessment of prognosis for patients. To enable fair comparison between the algorithms for this purpose, we set up the CAMELYON17 challenge in conjunction with the IEEE International Symposium on Biomedical Imaging 2017 Conference in Melbourne. Over 300 participants registered on the challenge website, of which 23 teams submitted a total of 37 algorithms before the initial deadline. Participants were provided with 899 whole-slide images (WSIs) for developing their algorithms. The developed algorithms were evaluated based on the test set encompassing 100 patients and 500 WSIs. The evaluation metric used was a quadratic weighted Cohen's kappa. We discuss the algorithmic details of the 10 best pre-conference and two post-conference submissions. All these participants used convolutional neural networks in combination with pre- and postprocessing steps. Algorithms differed mostly in neural network architecture, training strategy, and pre- and postprocessing methodology. Overall, the kappa metric ranged from 0.89 to -0.13 across all submissions. The best results were obtained with pre-trained architectures such as ResNet. Confusion matrix analysis revealed that all participants struggled with reliably identifying isolated tumor cells, the smallest type of metastasis, with detection rates below 40%. Qualitative inspection of the results of the top participants showed categories of false positives, such as nerves or contamination, which could be targeted for further optimization. Last, we show that simple combinations of the top algorithms result in higher kappa metric values than any algorithm individually, with 0.93 for the best combination.