Patch-wise 3D segmentation quality assessment combining reconstruction and regression networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) based semantic segmentation methods often struggle with 3D medical images due to complex structures and limited ground truth data.
• Accurate segmentation quality is critical for localized disease regions, not just global averages, for effective diagnosis.

Purpose of the Study:

• To develop a DL framework for predicting segmentation quality and identifying regions of inaccuracy in 3D medical images without requiring ground truth.
• To address the need for expeditious diagnosis by enabling reliable segmentation quality assessment.

Main Methods:

• A framework combining a 3D generative adversarial network (GAN) and a convolutional regression network was proposed.
• Conditional GAN reconstructs input images masked by segmentation results, and a regression network predicts patch-wise Dice Similarity Coefficient (DSC) based on segmentation.
• The method utilizes segmentation-derived features, eliminating the need for ground truth during inference.

Main Results:

• The method was evaluated on 3D knee MRI and lung CT datasets.
• Patch-wise DSC prediction achieved mean absolute errors of 0.01 for knee-MR and 0.04 for lung-CT.
• The framework successfully localized segmentation inaccuracies.

Conclusions:

• The proposed DL framework effectively identifies erroneous segmentation regions in 3D medical images.
• This capability can significantly aid downstream disease diagnosis and prognosis prediction.
• The method offers a promising approach for quality control in medical image segmentation.