Federated learning enables big data for rare cancer boundary detection

Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

Investigation of Asleep versus Awake Motor Mapping in Resective Brain Surgery

OBJECTIVE

To develop an asleep motor mapping paradigm for accurate detection of the corticospinal tract during glioma surgery and compare outcomes with awake patients undergoing glioma resection.

METHODS

A consecutive cohort of adult patients undergoing craniotomy for suspected diffuse glioma with tumor in a perirolandic location who had awake or asleep cortical and subcortical motor mapping with positive areas of motor stimulation were assessed for postoperative extent of resection (EOR), permanent neurological deficit, and proximity of stimulation to diffusion tensor imaging-based corticospinal tract depiction on preoperative magnetic resonance imaging. Outcome data were compared between asleep and awake groups.

RESULTS

In the asleep group, all 16 patients had improved or no change in motor function at last follow-up (minimum 3 months of follow-up). In the awake group, all 23 patients had improved function or no change at last follow-up. EOR was greater in the asleep group (mean [SD] EOR 88.71% [17.56%]) versus the awake group (mean [SD] EOR 80.62% [24.44%]), although this difference was not statistically significant (P = 0.3802). Linear regression comparing distance from stimulation to corticospinal tract in asleep (n = 14) and awake (n = 4) patients was r = -0.3759, R² = 0.1413, P = 0.1853, and 95% confidence interval = -0.4453 to 0.09611 and r = 0.7326, R² = 0.5367, P = 0.2674, and 95% confidence interval = -7.042 to 14.75, respectively.

CONCLUSION

In this small patient series, asleep motor mapping using commonly available motor evoked potential hardware appears to be safe and efficacious in regard to EOR and functional outcomes.

Differentiation of High-Grade Glioma and Intracranial Metastasis Using Volumetric Diffusion Tensor Imaging Tractography

OBJECTIVE

A reliable, noninvasive method to differentiate high-grade glioma (HGG) and intracranial metastasis (IM) has remained elusive. The aim of this study was to differentiate between HGG and IM using tumoral and peritumoral diffusion tensor imaging characteristics.

METHODS

A semiautomated script generated volumetric regions of interest (ROIs) for the tumor and a peritumoral shell at a predetermined voxel thickness. ROI differences in diffusion tensor imaging-related metrics between HGG and IM groups were estimated, including fractional anisotropy, mean diffusivity, total fiber tract counts, and tract density.

RESULTS

The HGG group (n = 46) had a significantly higher tumor-to-brain volume ratio than the IM group (n = 35) (P < 0.001). The HGG group exhibited significantly higher mean fractional anisotropy and significantly lower mean diffusivity within peritumoral ROI than the IM group (P < 0.05). The HGG group exhibited significantly higher total tract count and higher tract density in tumoral and peritumoral ROIs than the IM group (P < 0.05). Tumoral tract count and peritumoral tract density were the most optimal metrics to differentiate the groups based on receiver operating characteristic curve analysis. Predictive analysis using receiver operating characteristic curve thresholds was performed on 13 additional participants. Compared with correct clinical diagnoses, the 2 thresholds exhibited equal specificities (66.7%), but the tumoral tract count (85.7%) seemed more sensitive in differentiating the 2 groups.

CONCLUSIONS

Tract count and tract density were significantly different in tumoral and peritumoral regions between HGG and IM. Differences in microenvironmental interactions between the tumor types may cause these tract differences.

High-grade Gliomas Exhibit Higher Peritumoral Fractional Anisotropy and Lower Mean Diffusivity than Intracranial Metastases

Differentiating high-grade gliomas and intracranial metastases through non-invasive imaging has been challenging. Here, we retrospectively compared both intratumoral and peritumoral fractional anisotropy (FA), mean diffusivity (MD), and fluid-attenuated inversion recovery (FLAIR) measurements between high-grade gliomas and metastases. Two methods were utilized to select peritumoral region of interest (ROI). The first method utilized the manual placement of four ROIs adjacent to the lesion. The second method utilized a semiautomated and proprietary MATLAB script to generate an ROI encompassing the entire tumor. The average peritumoral FA, MD, and FLAIR values were determined within the ROIs for both methods. Forty patients with high-grade gliomas and 44 with metastases were enrolled in this study. Thirty-five patients with high-grade glioma and 30 patients with metastases had FLAIR images. There was no significant difference in age, gender, or race between the two patient groups. The high-grade gliomas had a significantly higher tumor-to-brain area ratio compared to the metastases. There were no differences in average intratumoral FA, MD, and FLAIR values between the two groups. Both the manual sample method and the semiautomated peritumoral ring method resulted in significantly higher peritumoral FA and significantly lower peritumoral MD in high-grade gliomas compared to metastases (p < 0.05). No significant difference was found in FLAIR values between the two groups peritumorally. Receiver operating curve analysis revealed FA to be a more sensitive and specific metric to differentiate high-grade gliomas and metastases than MD. The differences in the peritumoral FA and MD values between high-grade gliomas and metastases seemed due to the infiltration of glioma to the surrounding brain parenchyma.