Self-supervised-RCNN for medical image segmentation with limited data annotation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Deep Learning

Background:

• Supervised learning in medical imaging requires extensive expert annotations, which are costly and time-consuming.
• Limited labeled data poses a significant challenge for achieving high accuracy in medical image segmentation tasks.

Purpose of the Study:

• To propose a novel deep learning training strategy using self-supervised pretraining on unlabeled medical imaging data.
• To overcome the limitations of data scarcity in supervised medical image analysis.

Main Methods:

• A self-supervised pretraining approach was developed, involving applying arbitrary distortions to unlabeled images.
• A Mask-RCNN architecture was trained to detect and recover these distortions, enabling the model to learn image textures.
• The pretrained model was then fine-tuned for specific segmentation tasks using limited labeled data.

Main Results:

• The proposed self-supervised pretraining method demonstrated improved segmentation performance, with up to an 18% increase in the Dice score.
• The method showed effectiveness across various pretraining and fine-tuning scenarios, validated on the Osteoarthritis Initiative dataset for knee MRI effusion segmentation.

Conclusions:

• Self-supervised pretraining on unlabeled medical images provides a robust foundation for segmentation tasks with limited labeled data.
• This approach offers a valuable strategy for enhancing medical image analysis, applicable to anomaly detection, segmentation, and classification.