Detection and score grading for prostate adenocarcinoma using semantic segmentation.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Prostate cancer presents a significant global health burden.
• Accurate detection and grading are crucial for effective patient management.
• Distinguishing Gleason patterns 3 and 4 is critical for treatment decisions.

Purpose of the Study:

• To develop and evaluate a deep learning model for semantic segmentation of prostate adenocarcinoma.
• To improve the accuracy of prostate cancer detection and grading, specifically differentiating Gleason patterns 3 and 4.
• To provide a publicly available dataset for prostate cancer research.

Main Methods:

• Development of a novel dataset of 100 digitized whole-slide images of prostate needle core biopsies.
• Implementation of a deep learning model integrating dilated attention mechanisms and a residual convolutional U-Net architecture.
• Addressing class imbalance using pixel expansion and class weights, with five-fold cross-validation for robust evaluation.

Main Results:

• The model achieved an average Dice score of 0.87 and accuracy of 0.92 on internal cross-validation.
• On external test data, the model demonstrated an average Dice of 0.64 and accuracy of 0.81.
• Segmentation and grading results were validated by expert pathologists.

Conclusions:

• The proposed deep learning method shows potential for accurate prostate cancer detection and grading in clinical settings.
• The model's ability to distinguish Gleason patterns 3 and 4 offers a valuable tool for pathologists.
• The publicly available dataset supports further research and development in computational pathology.