Predicting PAM50 subtypes from whole slide images of prostate cancer biopsies

• Published in: Journal of clinical oncology Vol. 42; no. 16_suppl; p. e17001
• Main Authors: Ayad, Marina, Saft, Madeline, Li, Eric Victor, Kumar, Sai, Neill, Clayton, Patel, Hiten D., Schaeffer, Edward M., Liu, Yang, Davicioni, Elai, Yang, Ximing J, Cooper, Lee A.D., Ross, Ashley
• Format: Journal Article
• Language: English
• Published: American Society of Clinical Oncology 01.06.2024
• ISSN: 0732-183X; 1527-7755
• DOI: 10.1200/JCO.2024.42.16_suppl.e17001

e17001Background: Computational analysis of digitized pathological images has been shown to have diagnostic and prognostic applications. Recent research suggests that molecular features, previously defined at the genomic level, might be additionally recognized. In prostate cancer, transcriptomic PAM50 molecular subtypes may correlate with patient response to systemic therapy. Here we investigate whether computational analysis of digital prostate biopsy slides can predict PAM50 category. Methods: Our study utilized digitized images of 704 prostate biopsy slides from 336 patients with prostate cancer at Northwestern Memorial Hospital with Decipher testing (Veracyte, Inc San Diego, CA). All slides analyzed contain carcinoma with the majority being primary Gleason pattern 3 (545 out of 704). Hematoxylin and eosin stained slides were scanned on a Leica GT450 scanner at 40x magnification. PAM50 subtypes were derived from the Decipher Genomic Resource Information Database (GRID). Data were split at the subject level using 80% for training, 10% for validation, and 10% for evaluation, facilitating model development and assessment. We devised a model comprising two decoupled parts: a multi-instance learning model that utilizes a multi-class attention mechanism to explore relevant class-specific information from slides, and a pre-trained network to extract high-dimensional semantic features from the tiles of WSIs. The model trained based on the WSIs of the training set to predict PAM50 subtypes (luminal A, luminal B, and basal). This attention-based training approach allowed the model to capture relevant patterns and information present in the WSIs. Following the training phase, the model was subsequently evaluated on the testing set. Results: Model AUC on the evaluation set was 0.78. Among evaluation set misclassifications, no luminal samples were misclassified as basal subtype. Conclusions: Our investigation provides preliminary results on the ability to predict RNA expression-based subtypes from prostate cancer biopsy histology. This approach may help to preserve tissue, minimize costs, and decrease turnaround time associated with molecular testing in prostate cancer while offering patients with prostate cancer opportunities for precision medicine.