(P 138) A MULTI-INSTITUTIONAL ARCHIVE AND DEEP LEARNING PIPELINE FOR QUANTIFYING CHONDROSARCOMA COMPOSITION

Objective: Chondrosarcomas are often resistant to chemotherapy and radiation, and survival in the setting of metastatic disease is poor. Histologic grade guides management, but clinical decisions are impacted by high interobserver variability among pathologists and the limited prognostic power of grade alone. To address this, we developed a deep-learning pipeline that quantitatively describes chondrosarcoma composition on whole-slide images (WSIs).

Methods: We conducted a multicenter, retrospective study that included tumors from 230 patients. Patient survival by tumor grade and dedifferentiation status was evaluated with Kaplan–Meier (KM) analysis. Our computational workflow comprised the following steps: (1) partitioning each WSI into 225×225-pixel tiles; (2) background removal; (3) stain normalization; (4) tile feature extraction; (5) K-means clustering to identify visual phenotypes; (6) training a random-forest classifier to recognize those phenotypes; and (7) quantifying phenotype frequencies in new WSIs. Pipeline development used an initial subset of 22 WSIs from unique patients, and pipeline evaluation used 10 WSIs withheld during development.

Results: On KM analysis histologic grade and dedifferentiation status were significantly associated with overall survival. K-means clustering identified 10 visual phenotypes, and the random-forest model classified each with over 93% accuracy. Clusters represented normal bone elements (cluster 0 – bone, and cluster 3 –marrow) as well as five distinct tumor clusters (clusters 2, 4, 7, 8, and 9). Withheld WSIs were processed, tiles were classified, and a map of feature percentages was generated. Clusters with the highest incidence and average representation within this subset (clusters 4, 7) were tumor features.

Conclusion: Our results show that automated deep-learning analysis of WSIs may be able to objectively characterize chondrosarcoma histology. Future work will involve assessing the accuracy of the system relative to manual annotation, scaling the approach to the full cohort, refining feature extraction, and assessing correlations between computational phenotypes and clinical outcomes.