Knowledge distillation driven instance segmentation for grading prostate cancer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Prostate cancer (PCa) is a leading cause of cancer death in men, necessitating accurate early detection.
• Current deep learning models for PCa screening require extensive, well-annotated datasets, which are costly and time-consuming to acquire.
• This limitation can lead to suboptimal performance in clinical settings.

Purpose of the Study:

• To develop a novel knowledge distillation-based instance segmentation scheme for automated prostate cancer tissue extraction.
• To enable conventional semantic segmentation models to perform instance-aware segmentation using few-shot learning.
• To accurately grade prostate cancer using extracted tissue patterns and Gleason scores.

Main Methods:

• Implemented a knowledge distillation framework to enhance instance segmentation capabilities of semantic models.
• Utilized few-shot learning for incremental training on whole slide images (WSI).
• Extracted stroma, benign, and cancerous prostate tissues for Gleason score computation and PCa grading.

Main Results:

• The proposed scheme achieved superior performance in identifying prostate tissues, outperforming state-of-the-art methods by 2.01% and 4.45% in mean IoU on two large datasets.
• Achieved significant improvements in F1 scores for PCa grading (10.73% and 11.42%) compared to existing methods.
• Demonstrated strong agreement with expert pathologists in a blind study, with Pearson correlations of 0.9192 and 0.8984.

Conclusions:

• The novel knowledge distillation approach effectively enables instance segmentation for prostate cancer tissue analysis.
• The method significantly improves the accuracy of prostate cancer grading, aligning with clinical standards.
• This technique offers a resource-efficient and high-performing solution for automated prostate cancer screening and grading from WSI.