Masked Image Modeling Meets Self-Distillation: A Transformer-Based Prostate Gland Segmentation Framework for...

Area of Science:

• Digital pathology
• Machine learning in oncology
• Medical image analysis

Background:

• Accurate prostate cancer gland segmentation is crucial for grading and downstream machine learning tasks.
• Current methods for feature extraction are suboptimal due to reliance on cell-level or human-annotated data.
• Existing segmentation models have limitations for precise prostate cancer gland identification.

Purpose of the Study:

• To develop a reliable prostate gland segmentation framework for machine learning applications.
• To improve the accuracy of feature extraction for prostate cancer analysis.
• To address the limitations of current gland segmentation models.

Main Methods:

• Proposed a novel framework utilizing a dual-path Swin Transformer UNet architecture.
• Employed Masked Image Modeling for large-scale self-supervised pre-training on diverse data.
• Incorporated a tumor-guided self-distillation step to enhance encoder suitability for segmentation.

Main Results:

• Achieved state-of-the-art segmentation performance on two public datasets.
• Obtained a test mDice of 0.947 on the PANDA dataset.
• Achieved a test mDice of 0.664 on the SICAPv2 dataset.

Conclusions:

• The proposed framework demonstrates superior performance in prostate cancer gland segmentation.
• This advancement facilitates more accurate feature extraction for Gleason grading and survival analysis.
• The self-supervised approach with heterogeneous data integration enhances model generalizability.