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

Quality control (QC) in fetal MRI is essential for efficient, high-quality data acquisition and analysis aimed at assessing fetal brain development and detecting abnormalities. Supervised deep learning methods require numerous image-quality labels and have limited generalization to cross-domain out-of-distribution data. To address these problems, an Orientation Recognition Kolmogorov-Arnold Network (OR-KAN), composed of stacked Bottleneck KAN Convolution layers, was trained on turbo spin echo (TSE) T₂-weighted data augmented from seven fetal brain atlases to predict stack orientation (i.e., axial, sagittal, and coronal). Image quality was quantified as the entropy of the prediction uncertainty. Experiments showed that OR-KAN achieved an area under the receiver operating characteristic curve (AUROC) of 0.840 and an area under the precision recall curve (AUPR) of 0.954 on the TSE data. Its performance on balanced turbo field echo (BTFE) data, which exhibit a distinct T₂-weighted contrast, did not degrade (AUROC: 0.840 to 0.881; AUPR: 0.954 to 0.857). Moreover, bagging OR-KAN with models pre-trained on quality labels delivered the best results on both datasets, outperforming the state-of-the-art (SOTA) supervised method FetMRQC by 14.1 % on TSE (AUROC: 0.828 vs. 0.945) and by 21.9 % on BTFE (AUROC: 0.763 vs. 0.930). By selecting the highest-quality stack in each orientation, OR-KAN improved brain volume reconstruction success rates (with NiftyMIC) by 29.2 % (from 64.4 % to 83.2 %) and by 42.5 % (from 60.0 % to 85.5 %) for normal and abnormal fetuses, respectively. We anticipate its immediate utility in improving the diagnosis of fetal brain abnormalities and advancing the study of human brain development in utero across a wide range of clinical and neuroscientific applications. Code is available at: https://github.com/birthlab/OR-KAN.