Classification of acceleration and deceleration craniocerebral injuries by using cascaded deep learning models

Accurately determining the manner of craniocerebral injuries is critical in forensic practice, especially when distinguishing between acceleration and deceleration injuries. Traditional biomechanical methods often yield uncertain results due to atypical injury morphology. This study introduces a cascaded deep learning system that integrates a DeepLabv3 + segmentation network with a ResNet18 + ASPP classification network to automate injury mechanism analysis. An ablation study confirmed that segmentation-guided information significantly enhanced classification performance, with the macro-F1 score improving from 0.71 to 0.82. The segmentation network achieved a mean Dice coefficient of 0.87 for injury region delineation, and the cascaded model attained a mean AUC of 0.94 for injury mechanism classification. External validation across frontal, temporal, and occipital regions confirmed the model's generalizability, with performance patterns consistent with established biomechanical principles. The model also demonstrated high accuracy in distinguishing injury cases from normal images. This study offers new insights into objective injury mechanism determination for forensic applications.