Quality-label-free fetal brain MRI quality control based on image orientation recognition uncertainty

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Fetal MRI quality control (QC) is crucial for accurate assessment of fetal brain development and abnormality detection.
• Supervised deep learning methods for QC face challenges with data labeling and cross-domain generalization.

Purpose of the Study:

• To develop a novel deep learning model for automated fetal MRI quality control.
• To improve the generalization capability of QC methods across different MRI sequences.

Main Methods:

• An Orientation Recognition Kolmogorov-Arnold Network (OR-KAN) was developed using stacked Bottleneck KAN Convolution layers.
• The OR-KAN was trained on augmented fetal brain atlases to predict image orientation (axial, sagittal, coronal).
• Image quality was quantified using prediction uncertainty entropy.

Main Results:

• OR-KAN achieved high performance on T2-weighted TSE data (AUROC: 0.840, AUPR: 0.954) and maintained performance on BTFE data (AUROC: 0.881, AUPR: 0.857).
• Bagging OR-KAN with pre-trained models outperformed the SOTA supervised method FetMRQC by up to 21.9%.
• OR-KAN improved fetal brain volume reconstruction success rates by up to 42.5% for both normal and abnormal fetuses.

Conclusions:

• OR-KAN offers a robust and generalizable solution for fetal MRI quality control.
• The method significantly enhances the accuracy of fetal brain volume reconstruction.
• This advancement is expected to improve the diagnosis of fetal brain abnormalities and in utero brain development studies.