A Weak and Semi-supervised Segmentation Method for Prostate Cancer in TRUS Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer segmentation in ultrasound images is crucial for diagnosis and treatment planning.
• Manual segmentation by radiologists is time-consuming and requires extensive expertise.
• Deep learning models typically require large, fully annotated datasets, which are often unavailable.

Purpose of the Study:

• To develop and evaluate a weak and semi-supervised deep learning framework for automated prostate cancer segmentation.
• To reduce the reliance on fully annotated data for training segmentation models.
• To alleviate the burden on radiologists by automating the segmentation process.

Main Methods:

• A novel weak and semi-supervised deep learning framework was proposed.
• The framework utilized a combination of strongly (pixel-wise) and weakly (lesion location) supervised data.
• Iterative retraining and label refinement were employed to improve segmentation accuracy using limited annotations.

Main Results:

• The proposed method achieved a mean intersection over union (mIoU) of approximately 0.6 with 40% strong supervision.
• Performance was comparable to, though slightly lower than, fully supervised methods (approx. 2% decrease).
• The framework demonstrated effectiveness in training neural networks with incomplete annotations.

Conclusions:

• Weak and semi-supervised learning offers a viable approach for prostate cancer segmentation in ultrasound images.
• The developed framework can significantly reduce the need for extensive manual annotation.
• This method holds promise for improving efficiency and accessibility in radiological workflows.