QCResUNet: Joint subject-level and voxel-level segmentation quality prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Deep learning models excel at brain tumor segmentation from MRI scans.
• Segmentation reliability is hindered by outliers and out-of-distribution samples, limiting clinical use.
• Existing quality control methods are often limited to single-modality cardiac MRI and lack voxel-level error identification.

Purpose of the Study:

• To develop a novel multi-task deep learning architecture, QCResUNet, for automated segmentation quality control.
• To provide both subject-level quality measures and voxel-level error maps for segmentation refinement.
• To address limitations of prior methods by handling multi-class segmentation and identifying specific erroneous regions.

Main Methods:

• Proposed QCResUNet, a multi-task deep learning model for segmentation quality assessment.
• Validated on brain tumor segmentation tasks using internal (BraTS 2021) and external datasets (BraTS-SSA, WUSM).
• Evaluated performance on cardiac MRI segmentation from the Automated Cardiac Diagnosis Challenge (ACDC).

Main Results:

• QCResUNet achieved high performance in predicting subject-level segmentation quality.
• The model accurately identified segmentation errors on a voxel basis across different datasets.
• Demonstrated effectiveness in assessing segmentation quality for both brain tumor and cardiac MRI data.

Conclusions:

• QCResUNet offers a robust solution for automated quality control in medical image segmentation.
• The voxel-level error maps can guide refinement, facilitating human-in-the-loop feedback for improved clinical segmentations.
• This method enhances the reliability and clinical applicability of deep learning-based segmentation tools.