Learning anatomy from unlabelled CT volumes: A self-supervised framework for improving prostate radiotherapy...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate CT-based structure segmentation is vital for prostate cancer radiotherapy.
• Manual contouring of organs like the prostate, seminal vesicles (SV), and penile bulb (PB) is time-consuming and variable.
• Deep learning models require large labeled datasets, which are often scarce in clinical practice.

Purpose of the Study:

• To evaluate if self-supervised (label-free) slice-prediction pretraining can improve segmentation performance, particularly with limited annotated data.
• To assess the impact of self-supervised pretraining on the accuracy of pelvic structure segmentation in CT images.

Main Methods:

• Utilized 322 pelvic CT volumes for training and testing.
• Pretrained a lightweight 2D U-Net encoder on unlabeled data using a slice-prediction task.
• Fine-tuned the model for multi-class segmentation using full or reduced (60 patients) labeled datasets.
• Assessed segmentation accuracy using Mean Distance Agreement (MDA) and Dice Similarity Coefficient (DSC).

Main Results:

• Self-supervised pretraining consistently reduced MDA across major pelvic structures.
• Significant MDA reductions observed for bladder (0.600mm to 0.547mm), prostate (1.281mm to 1.183mm), and SV (1.175mm to 0.893mm).
• Improvements were noted even with limited annotated data (60 patients).

Conclusions:

• Self-supervised pretraining via slice prediction facilitates anatomically informed feature learning.
• This approach enhances segmentation robustness and accuracy under limited data conditions.
• The method is label-free during pretraining and compatible with lightweight architectures, suiting resource-constrained environments.