Optimizing MRI sequence classification performance: insights from domain shift analysis.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Deep learning for medical diagnostics

Background:

• Multicenter MRI studies face challenges in sequence classification due to protocol variability and domain shift, impacting automated model accuracy.
• Existing automated MRI sequence identification models struggle with domain shift, particularly between adult and pediatric datasets.
• Manual annotation is labor-intensive, highlighting the need for robust automated solutions.

Purpose of the Study:

• To evaluate the effectiveness of pre-trained deep learning models in addressing domain shift in MRI sequence classification.
• To compare the performance of a convolutional neural network (ResNet) and a CNN-Transformer hybrid model (MedViT) on adult-to-pediatric MRI data.
• To investigate the impact of expert domain knowledge adjustments on model accuracy for pediatric MRI data.

Main Methods:

• Retrospective, multicentric study utilizing adult MRI data for training and pediatric MRI data for testing.
• Employed pre-trained ResNet-18 and MedViT models, a hybrid CNN-Transformer architecture.
• Applied expert domain knowledge adjustments to account for differences in MRI sequence types between adult and pediatric datasets.

Main Results:

• The MedViT model achieved higher accuracy (0.893) than ResNet-18 and benchmark models in classifying pediatric MRI sequences.
• Expert domain knowledge adjustments further enhanced MedViT's accuracy to 0.905, demonstrating improved robustness.
• The findings indicate MedViT's superior capability in handling domain shift from adult to pediatric MRI data.

Conclusions:

• Advanced neural network architectures, such as MedViT, are crucial for robust MRI sequence classification under domain shift.
• Integrating expert domain knowledge with deep learning models significantly improves accuracy in diverse datasets.
• Hybrid architectures combining CNNs and transformers offer enhanced reliability for automated MRI sequence classification in multicenter research and clinical practice.