An Infant Brain MR Image Segmentation Method with Attention Mechanism, Multi-Resolution Deep Supervision and Class Weight Tuning Strategy.

Background: Infant brain tissue segmentation is critical for understanding early brain development and detecting neurodevelopmental abnormalities. However, accurate segmentation remains challenging due to partial-volume effects, imaging noise, and low contrast between tissues in isointense regions. Objective: This study aimed to automatically segment infant brain Magnetic Resonance Imaging (MRI) images into white matter, gray matter, and cerebrospinal fluid. Material and Methods: In this analytical study, we addressed the challenges of infant brain tissue segmentation. We proposed a deep, lightweight dense backbone architecture incorporating transition blocks. The backbone was enhanced with a spatial-channel attention mechanism to capture global dependencies and emphasize task-relevant features, as well as multi-resolution deep supervision to mitigate vanishing gradients and reduce overfitting. Moreover, we used class weighting strategy to deal with class imbalance and low contrast between brain tissues. Results: Our method attains Dice similarity scores of 91.80%, 92.07% and 94.58% for white matter, gray matter, and cerebrospinal fluid, respectively. Quantitative and qualitative evaluations show that the proposed approach accurately segments infant brain MR images and reduces errors in isointense regions. Conclusion: The results demonstrate that the proposed method precisely delineates infant brain tissues and effectively overcomes common segmentation challenges.