Deep learning for detection and automatic visualization of radiation-induced temporal lobe injury in nasopharyngeal carcinoma across endemic and non-endemic areas in China.

PURPOSE: Detection of radiation-induced temporal lobe injury (RTLI) at the earliest radiologically detectable stage is important for timely intervention in nasopharyngeal carcinoma but remains challenging due to subtle MRI findings. This study aimed to develop and validate an MRI-based multi-task deep learning (MTL) model for RTLI detection and automated lesion visualization across endemic and non-endemic regions.

METHODS

A total of 956 patients with nasopharyngeal carcinoma were retrospectively and prospectively enrolled from southern and northern China. Axial T2-weighted MRI was used as model input. A 2.5D ResNet-based MTL network integrating RTLI classification and segmentation was developed and evaluated in internal, external and prospective test cohorts. Diagnostic performance was assessed using area under the receiver operating characteristic curve (AUC) and Dice similarity coefficient (DSC). A multi-reader study assessed the impact of MTL assistance across readers with varying experience levels, and gradient-weighted class activation mapping was used to visualize model attention.

RESULTS

The MTL model achieved AUCs of 0.974, 0.969, and 0.953 in the validation, southern internal test, and northern external test cohorts, respectively, with sensitivities of 0.914, 0.899, and 0.854. Automated RTLI segmentation achieved DSCs of 0.720, 0.690, and 0.711. It showed sensitivity comparable to expert reader and outperformed competent reader. In the retrospective reader study, MTL assistance improved diagnostic performance for the evaluated readers, with prospective confirmation of clinical benefit pending.

CONCLUSIONS

This MRI-based MTL framework enables accurate detection and visualization of RTLI during follow-up, improves diagnostic performance for evaluated readers, and shows potential for clinical application across regions.