OS03.6.A UNSUPERVISED CLUSTERING OF MORPHOLOGY PATTERNS ON WHOLE SLIDE IMAGES GUIDE PROGNOSTIC STRATIFICATION OF GLIOBLASTOMA PATIENTS

Abstract

Abstract BACKGROUND Glioblastoma is the most common malignant adult brain tumor with a poor prognosis and heterogeneous morphology. Stratifying glioblastoma patients according to overall survival (OS) from H&E-stained histopathology whole slide images (WSI) using advanced computational methods is a challenging task with direct clinical implications. We hypothesize that quantifying morphology patterns present in WSI can yield biomarkers of prognostic relevance contributing to optimizing clinical decision-making. MATERIAL AND METHODS This work is based on 188 glioblastoma (IDH-wt, grade 4) cases identified in the TCGA-GBM and TCGA-LGG, based on the 2021 WHO classification criteria. One Diagnostic WSI from each patient at 10X magnification was selected (ntraining/nhold-out-test =152 /36) and labeled as short (<9 months, n=94) or long (>13 months, n=94) survivors. The WSI is split into non-overlapping 256x256 patches, and extensive patch-level WSI curation is conducted to discard artifactual content, i.e., glass reflections, pen markings, black lines on the slide, and tissue tearing. Each patch is described by a feature vector of 512 dimensions using a pre-trained VGG16. Principal component analysis reduced the dimensionality of these vectors to 236 such that 85% variance is retained. Unsupervised k-means clustering revealed distinct groups of morphology patterns, where the number of clusters is automatically determined based on the rand index and silhouette coefficient. The proportions of these patterns describe the tumor’s spatial heterogeneity and are used to distinguish between short and long survivors using a random forest classifier. RESULTS We identified seven clusters of distinct morphology patterns, including 3 categories of tumor cellularity (low/intermediate/high), necrosis/macrophages, and other prognostically-relevant characteristics. The patient-level short vs long survivor prediction is driven by the proportion of these patterns present in WSI achieving classification accuracy, sensitivity and specificity of 83.83%, 94.44%, and 72.72% respectively on the hold-out test set. The high sensitivity signifies the algorithm’s ability to accurately predict the short survivors which is important in treatment planning and patient management. CONCLUSION The interpretability analysis of various morphologic patterns in distinct histologic sub-regions of glioblastoma may uncover correlations with short and long survivors, providing a deeper understanding of the complex relationship between histologic features and clinical outcomes and shedding light on the heterogeneity of glioblastoma. This association could offer additional prognostic information to clinical neuropathologists during microscopic assessment, leading to more refined prognostication and potentially impacting clinical decision-making for improved patient outcomes.