Joint Segment-Level and Pixel-Wise Losses for Deep Learning Based Retinal Vessel Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Current deep learning models for retinal vessel segmentation use pixel-wise losses.
• This approach equally weights all vessel pixels, hindering the accurate segmentation of thin vessels due to imbalanced pixel ratios in fundus images.
• Accurate segmentation of thin vessels is crucial for diagnosing eye diseases.

Purpose of the Study:

• To propose a new segment-level loss function to improve the segmentation of thin retinal vessels.
• To enhance deep learning models for retinal vessel segmentation by balancing the importance of thick and thin vessels.

Main Methods:

• Developed a novel segment-level loss function that emphasizes thickness consistency of thin vessels.
• Combined the proposed segment-level loss with traditional pixel-wise losses in a joint-loss framework.
• Trained and evaluated deep learning models using the joint-loss framework on public datasets.

Main Results:

• The joint-loss framework significantly outperformed state-of-the-art methods in both separate-training and cross-training evaluations.
• Models trained with joint losses learned more distinguishable features for vessel segmentation compared to pixel-wise loss alone.
• Consistent performance improvements were observed across both deep and shallow network architectures.

Conclusions:

• The proposed joint-loss framework effectively addresses the limitations of pixel-wise losses in retinal vessel segmentation.
• The segment-level loss component is key to improving the accuracy of thin vessel segmentation.
• This approach offers a versatile method for enhancing deep learning models without altering network architectures.