A multi-task segFormer framework for lesion segmentation and cerebral palsy classification based on multi-modal MRI in infant with periventricular white matter injury

Objective: This study aimed to develop a multi-task framework for detection of five MRI predictors of CP and intelligent recognition of CP based on multi-modal MRI in infant with PVWMI. Methods: We present MMSeg-CP, a multi-task framework for joint anatomical target region segmentation, lesion of target region segmentation, and CP classification from registered T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI). MMSeg-CP adopts a SegFormer-based hierarchical transformer encoder and a lightweight all-MLP decoder, followed by lesion prediction and AttentionPool2d-based classification heads for infant neuroimaging characteristics, and performance was evaluated through five-fold cross-validation against nine comparative architectures using overlap, boundary, and classification metrics. Results: The study included 122 PVWMI infants (90 PVWMI with CP and 32 PVWMI with non-CP) and 121 infants with normal MRI. In five-fold cross-validation, the model achieved mean Dice values of 0.79 for target regions and 0.41 for lesions of target region, along with 0.95 slice-level accuracy and 0.88 subject-level accuracy. Compared with nine representative baseline models, MMSeg-CP provided the best overall balance between overlap accuracy, boundary precision, specificity, and sensitivity. Conclusion: MMSeg-CP enables joint detection of five MRI predictors of CP and intelligent CP recognition, supporting its potential as a clinical decision-support tool for early CP screening.