Using Deep Learning for Whole Slide Image Prostate Cancer Diagnosis and Grading in South Florida Veteran Population

Abstract

Listen

Translate article

Abstract Introduction/Objective Prostate cancer is the most common non-cutaneous malignancy in veterans, with approximately 11,000 new prostate cancer cases diagnosed in the Veteran’s Affairs system each year. Prostate cancer diagnosis and grading can be challenging even for experienced pathologists. Although large VA medical centers have pathologists that specialize in urologic pathology, the vast majority have not. We hypothesized that the AI-augmented diagnosis and grading may provide the solution for such situations. Methods/Case Report Dataset consisted of 10,000 prostate biopsy whole slide images (WSI) from the Kaggle PANDA challenge, and 6,000 WSI from the James A. Haley Veterans’ Hospital. Two Classification models were trained on the combined Kaggle and VA datasets using whole slide labels, and not annotated slides that resemble semi-supervised training. Two-Class Classification to predict Benign: ISUP [0] / Cancerous: ISUP [1,2,3,4,5] Three-Class Classification to predict Benign: ISUP [0] / Low-grade: ISUP [1,2] / High-grade: ISUP [3,4,5] WSI split into “tiles” were used for training the models to reduce whitespace around samples, manage large images, and normalize dimensions/orientations. Results (if a Case Study enter NA) Models trained purely as binary and 3-class classifiers performed very well. Two-Class Model: Accuracy = 0.937 Precision = 0.965 F1 = 0.94 AUC = 0.979 Three-Class Model: Accuracy 0.89 o Benign: Precision=0.897 f1=0.928 o Low-grade: Precision=0.866 f1=0.841 o High-grade: Precision=0.91 f1=0.878 We plan to develop multi-stage prediction models using these 2-Class and 3-Class classifiers as the first stage and a cancer grade predictor in the second stage. Conclusion We successfully showed that AI can augment pathologist’s diagnosis and grading of prostate cancer.