SegQC: a segmentation network-based framework for multi-metric segmentation quality control and segmentation error...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of structures in volumetric medical images is crucial for clinical practice and AI model development.
• Existing quality control (QC) methods have limitations in accuracy and error detection.

Purpose of the Study:

• Introduce SegQC, a novel framework for segmentation quality estimation and error detection in volumetric medical images.
• Evaluate SegQC's performance against unsupervised Test Time Augmentation (TTA) and supervised autoencoder (AE) based QC methods.

Main Methods:

• SegQCNet: a deep network for voxel-wise segmentation error probability estimation.
• Three novel segmentation quality metrics derived from error probabilities.
• A new method for detecting segmentation errors within slices.
• Evaluation using expert radiologist corrections on fetal MRI scans.

Main Results:

• SegQC outperforms TTA-based QC in quality estimation for fetal brain and body segmentation.
• SegQC achieves lower Mean Absolute Error (MAE) and higher Pearson correlation for Dice estimates compared to TTA and AE methods.
• Segmentation error detection achieved high recall and precision rates for fetal structures.
• SegQC's ranking metrics surpass TTA and SegQCNet estimations.

Conclusions:

• SegQC offers high-quality metrics estimation for 2D and 3D medical images.
• The framework provides effective error localization within slices, significantly improving segmentation QC.
• SegQC demonstrates superior performance in both quality estimation and error detection compared to existing methods.