A multiparametric MRI deep learning framework for automated segmentation and four-category classification of solitary spinal lesions
Related Experiment Videos
Jun Xu1, Chenxi Wang1, Wenjia Wang2
1Department of Radiology, Peking University Third Hospital, No. 49, North Garden Road, Haidian District, Beijing, 100069, China.
Background:
Accurate preoperative categorization of solitary spinal lesions (SSLs) is important because management differ across lesion categories.
Methods:
This retrospective study included 600 patients with SSLs (332 men, 268 women; median age, 53 years) examined from January 2018 to March 2025 and assigned to training, validation, and test sets using a fixed temporal split. According to histopathology, lesions were categorized as malignant, intermediate, benign, or non-neoplastic. Lesions on CE-T1WI and DWI (b = 50 s/mm²) images from multiple-b-value diffusion-weighted imaging (mb-DWI) were segmented using 3D nnU-Net. Two DenseNet-121 models were developed: a structural MRI model (DL_struct) based on co-registered CE-T1WI, T2WI, and T2-FS images, and a mb-DWI model (DL_mb) based on images acquired at b values of 50, 400, and 800 s/mm² and the corresponding ADC maps. Their late fusion yielded an imaging model (DL_fusion). A clinical model based on five clinico-radiological features was built and integrated with DL_fusion to generate the combined model. Performance was compared with five radiologists.
Results:
In the test cohort, Dice scores were 0.900 for CE-T1WI and 0.857 for DWI (b = 50 s/mm²). The combined model achieved the best four-category classification performance, with a macro-AUC/accuracy of 0.921/0.772, compared with 0.916/0.743 for DL_fusion, 0.852/0.640 for DL_struct, 0.871/0.647 for DL_mb, and 0.721/0.493 for the clinical model. Category-specific AUCs ranged from 0.893 to 0.956. The combined model achieved an accuracy of 0.772, outperforming the junior and attending radiologists (0.588-0.750), although it remained inferior to the senior radiologist (0.853). Model assistance improved reader performance, with the largest gains in junior radiologists.
Conclusion:
This multimodal DL framework enabled automated segmentation and four-category classification of SSLs and improved the performance of less experienced readers.