SESV: Accurate Medical Image Segmentation by Predicting and Correcting Errors

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Deep convolutional neural networks (DCNNs) are vital for medical image segmentation but require further accuracy and robustness for clinical use.
• Existing DCNNs often produce segmentation errors that limit their application in diagnosis.

Purpose of the Study:

• To propose a novel and generic framework, Segmentation-Emendation-reSegmentation-Verification (SESV), to improve the accuracy of existing DCNNs for medical image segmentation.
• To enhance segmentation performance without designing entirely new segmentation models.

Main Methods:

• The SESV framework predicts segmentation errors and uses these predictions to refine segmentation masks.
• It incorporates a re-segmentation network that uses error maps as priors and a verification network for region-based mask acceptance/rejection.

Main Results:

• The SESV framework significantly improved the accuracy of DeepLabv3+ in segmenting gland cells, skin lesions, and retinal microaneurysms across CRAG, ISIC, and IDRiD datasets.
• Consistent improvements were observed when applying SESV with other DCNNs like PSPNet, U-Net, and FPN.

Conclusions:

• The proposed SESV framework effectively enhances the accuracy of various DCNNs for diverse medical image segmentation tasks.
• SESV offers a versatile approach to improve clinical applicability of existing deep learning segmentation models.