Classifying Malignancy in Prostate Glandular Structures from Biopsy Scans with Deep Learning.

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Histopathological classification of prostate cancer, particularly Gleason grading, is subjective and relies heavily on expert pathologists.
• Accurate Gleason grading is critical for determining prognosis and guiding treatment strategies for prostate cancer patients.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated identification of Gleason patterns in prostate cancer histology images.
• To assess the performance of DL models in discriminating between benign tissue, Gleason score 3 (GS3), and Gleason score 4 (GS4) prostate cancer.

Main Methods:

• Utilized a curated cohort of prostate cancer histology images, partitioned into 14,509 tiles, with expert-assigned Gleason patterns.
• Employed transfer learning and fine-tuning of deep neural network architectures, pre-trained on ImageNet and adapted for histopathological analysis.
• Compared various DL network architectures for their efficacy in classifying cancer grades.

Main Results:

• The best-performing DL model achieved 91% accuracy, 0.91 F1-score, and 0.96 AUC in discriminating cancer from benign tissue (52 patients).
• The model demonstrated moderate performance in differentiating GS3 from GS4 prostate cancer, with 68% accuracy and 0.71 AUC (40 patients).

Conclusions:

• Deep learning models show significant potential for assisting pathologists in prostate cancer histopathological classification.
• Further refinement of DL algorithms is needed to improve the accuracy of distinguishing between specific Gleason patterns (GS3 vs. GS4).