Automated error localisation and correction techniques for deep-learning-based segmentation of 3D MRI sequences based...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Convolutional neural networks (CNNs) automate musculoskeletal imaging segmentation, reducing processing time.
• Manual correction is often required due to reliability concerns, increasing workload.

Purpose of the Study:

• To develop a method leveraging network-derived uncertainty for automatic identification and localization of segmentation errors in hip MRI.
• To reduce the need for exhaustive manual review of CNN-based segmentations.

Main Methods:

• Trained a 3D nnU-Net on delayed gadolinium-enhanced MRI of hip cartilage.
• Computed voxel-wise uncertainty scores from ensembled sub-networks.
• Aggregated uncertainty over supervoxels and evaluated impact on clinical metrics to generate risk scores for targeted correction.

Main Results:

• Guided supervoxel correction using risk scores reduced mean absolute relative error by 2.1-fold.
• Guided manual correction achieved a 3.5-fold reduction, demonstrating ~62% supervoxel correction efficiency.
• Correcting the top 10 regions yielded up to 88% efficiency, highlighting targeted correction benefits.

Conclusions:

• This approach serves as a proof-of-concept for targeted correction in hip MRI.
• 3D feature-derived uncertainty aggregation can significantly reduce the correction burden compared to traditional methods.
• Enhances the clinical utility of CNN-based segmentation in medical imaging.