Predicting local persistence/recurrence after radiation therapy for head and neck cancer from PET/CT using a multi-objective, multi-classifier radiomics model

Do We Need to Add the Type of Treatment Planning System, Dose Calculation Grid Size, and CT Density Curve to Predictive Models?

Background: Generalizability and domain dependency are critical challenges in developing predictive models for healthcare, particularly in medical diagnostics and radiation oncology. Predictive models designed to assess tumor recurrence rely on comprehensive and high-quality datasets, encompassing treatment planning parameters, imaging protocols, and patient-specific data. However, domain dependency, arising from variations in dose calculation algorithms, computed tomography (CT) density conversion curves, imaging modalities, and institutional protocols, can significantly undermine model reliability and clinical utility. Methods: This study evaluated dose calculation differences in the head and neck cancer treatment plans of 19 patients using two treatment planning systems, Pinnacle 9.10 and RayStation 11, with similar dose calculation algorithms. Variations in the dose grid size and CT density conversion curves were assessed for their impact on domain dependency. Results: Results showed that dose grid size differences had a more significant influence within RayStation than Pinnacle, while CT curve variations introduced potential domain discrepancies. The findings underscore the critical role of precise and standardized treatment planning in enhancing the reliability of predictive modeling for tumor recurrence assessment. Conclusions: Incorporating treatment planning parameters, such as dose distribution and target volumes, as explicit features in model training can mitigate the impact of domain dependency and enhance prediction accuracy. Solutions such as multi-institutional data harmonization and domain adaptation techniques are essential to improve model generalizability and robustness. These strategies support the better integration of predictive modeling into clinical workflows, ultimately optimizing patient outcomes and personalized treatment strategies.

A comparison of different machine learning classifiers in predicting xerostomia and sticky saliva due to head and neck radiotherapy using a multi-objective, multimodal radiomics model

Background and Purpose. Although radiotherapy techniques are a primary treatment for head and neck cancer (HNC), they are still associated with substantial toxicity and side effects. Machine learning (ML) based radiomics models for predicting toxicity mostly rely on features extracted from pre-treatment imaging data. This study aims to compare different models in predicting radiation-induced xerostomia and sticky saliva in both early and late stages HNC patients using CT and MRI image features along with demographics and dosimetric information.Materials and Methods.A cohort of 85 HNC patients who underwent radiation treatment was evaluated. We built different ML-based classifiers to build a multi-objective, multimodal radiomics model by extracting 346 different features from patient data. The models were trained and tested for prediction, utilizing Relief feature selection method and eight classifiers consisting eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), Naive Bayes (NB), Logistic Regression (LR), and Decision Tree (DT). The performance of the models was evaluated using sensitivity, specificity, area under the curve (AUC), and accuracy metrics.Results.Using a combination of demographics, dosimetric, and image features, the SVM model obtained the best performance with AUC of 0.77 and 0.81 for predicting early sticky saliva and xerostomia, respectively. Also, SVM and MLP classifiers achieved a noteworthy AUC of 0.85 and 0.64 for predicting late sticky saliva and xerostomia, respectively.Conclusion. This study highlights the potential of baseline CT and MRI image features, combined with dosimetric data and patient demographics, to predict radiation-induced xerostomia and sticky saliva. The use of ML techniques provides valuable insights for personalized treatment planning to mitigate toxicity effects during radiation therapy for HNC patients.

Radiomics Features on Enhanced Computed Tomography Predict FOXP3 Expression and Clinical Prognosis in Patients with Head and Neck Squamous Cell Carcinoma

Forkhead box P3 (FOXP3) has been identified as a novel molecular marker in various types of cancer. The present study assessed the expression of FOXP3 in patients with head and neck squamous cell carcinoma (HNSCC) and its potential as a clinical prognostic indicator, and developed a radiomics model based on enhanced computed tomography (CT) imaging. Data from 483 patients with HNSCC were downloaded from the Cancer Genome Atlas for FOXP3 prognostic analysis and enhanced CT images from 139 patients included in the Cancer Imaging Archives, which were subjected to the maximum relevance and minimum redundancy and recursive feature elimination algorithms for radiomics feature extraction and processing. Logistic regression was used to build a model for predicting FOXP3 expression. A prognostic scoring system for radiomics score (RS), FOXP3, and patient clinicopathological factors was established to predict patient survival. The area under the receiver operating characteristic (ROC) curve (AUC) and calibration curve and decision curve analysis (DCA) were used to evaluate model performance. Furthermore, the relationship between FOXP3 and the immune microenvironment, as well as the association between RS and immune checkpoint-related genes, was analyzed. Results of analysis revealed that patients with HNSCC and high FOXP3 mRNA expression exhibited better overall survival. Immune infiltration analysis revealed that FOXP3 had a positive correlation with CD4 + and CD8 + T cells and other immune cells. The 8 best radiomics features were selected to construct the radiomics model. In the FOXP3 expression prediction model, the AUC values were 0.707 and 0.702 for the training and validation sets, respectively. Additionally, the calibration curve and DCA demonstrated the positive diagnostic utility of the model. RS was correlated with immune checkpoint-related genes such as ICOS, CTLA4, and PDCD1. A predictive nomogram was established, the AUCs were 0.87, 0.787, and 0.801 at 12, 24, and 36 months, respectively, and DCA demonstrated the high clinical applicability of the nomogram. The enhanced CT radiomics model can predict expression of FOXP3 and prognosis in patients with HNSCC. As such, FOXP3 may be used as a novel prognostic marker to improve individualized clinical diagnosis and treatment decisions.

Methodological evaluation of original articles on radiomics and machine learning for outcome prediction based on positron emission tomography (PET)

AIM

Despite a vast number of articles on radiomics and machine learning in positron emission tomography (PET) imaging, clinical applicability remains limited, partly owing to poor methodological quality. We therefore systematically investigated the methodology described in publications on radiomics and machine learning for PET-based outcome prediction.

METHODS

A systematic search for original articles was run on PubMed. All articles were rated according to 17 criteria proposed by the authors. Criteria with >2 rating categories were binarized into "adequate" or "inadequate". The association between the number of "adequate" criteria per article and the date of publication was examined.

RESULTS

One hundred articles were identified (published between 07/2017 and 09/2023). The median proportion of articles per criterion that were rated "adequate" was 65% (range: 23-98%). Nineteen articles (19%) mentioned neither a test cohort nor cross-validation to separate training from testing. The median number of criteria with an "adequate" rating per article was 12.5 out of 17 (range, 4-17), and this did not increase with later dates of publication (Spearman's rho, 0.094; p = 0.35). In 22 articles (22%), less than half of the items were rated "adequate". Only 8% of articles published the source code, and 10% made the dataset openly available.

CONCLUSION

Among the articles investigated, methodological weaknesses have been identified, and the degree of compliance with recommendations on methodological quality and reporting shows potential for improvement. Better adherence to established guidelines could increase the clinical significance of radiomics and machine learning for PET-based outcome prediction and finally lead to the widespread use in routine clinical practice.

Performance Analysis of Six Semi-Automated Tumour Delineation Methods on [18F] Fluorodeoxyglucose Positron Emission Tomography/Computed Tomography (FDG PET/CT) in Patients with Head and Neck Cancer

Background. Head and neck cancer (HNC) is the seventh most common neoplastic disorder at the global level. Contouring HNC lesions on [18F] Fluorodeoxyglucose positron emission tomography/computed tomography (FDG PET/CT) scans plays a fundamental role for diagnosis, risk assessment, radiotherapy planning and post-treatment evaluation. However, manual contouring is a lengthy and tedious procedure which requires significant effort from the clinician. Methods. We evaluated the performance of six hand-crafted, training-free methods (four threshold-based, two algorithm-based) for the semi-automated delineation of HNC lesions on FDG PET/CT. This study was carried out on a single-centre population of n=103 subjects, and the standard of reference was manual segmentation generated by nuclear medicine specialists. Figures of merit were the Sørensen-Dice coefficient (DSC) and relative volume difference (RVD). Results. Median DSC ranged between 0.595 and 0.792, median RVD between -22.0% and 87.4%. Click and draw and Nestle's methods achieved the best segmentation accuracy (median DSC, respectively, 0.792 ± 0.178 and 0.762 ± 0.107; median RVD, respectively, -21.6% ± 1270.8% and -32.7% ± 40.0%) and outperformed the other methods by a significant margin. Nestle's method also resulted in a lower dispersion of the data, hence showing stronger inter-patient stability. The accuracy of the two best methods was in agreement with the most recent state-of-the art results. Conclusions. Semi-automated PET delineation methods show potential to assist clinicians in the segmentation of HNC lesions on FDG PET/CT images, although manual refinement may sometimes be needed to obtain clinically acceptable ROIs.

Surveillance of head neck cancer: Case for personalized and standardized surveillance

Although surgery, radiotherapy, chemotherapy, or combined treatment often elicits an initial satisfactory response, relapses are frequently observed within two years. Current surveillance methods, including clinical exams and imaging evaluations, have not unambiguously demonstrated a survival benefit, most probably due to a lack of sensitivity in detecting very early recurrence. Current guidelines advise post-treatment surveillance of head and neck cancer (HNC) patients should involve scheduled appointments with a variety of practitioners. The benefits of prolonged routine follow-up on survival have not been proven. Increasing numbers of HNC survivors raise the burden to provide efficient and effective care.