[Multicenter study of deep learning multitask model for predicting neoadjuvant chemotherapy response and event-free...

Objective: To investigate the value of a pretreatment magnetic resonance imaging (MRI)-based deep learning multitask model in the simultaneous prediction of neoadjuvant chemotherapy response (NACr) and event-free survival (EFS) in children with head and neck rhabdomyosarcoma (RMS). Methods: A total of 382 children with stage III/IV head and neck RMS according to the Intergroup Rhabdomyosarcoma Study Group (IRS) staging system were retrospectively enrolled from Beijing Children's Hospital, Beijing Tongren Hospital, and Shunyi Women and Children's Hospital of Beijing Children's Hospital. The internal cohort from Beijing Children's Hospital (n=322) was randomly divided into a training set (n=257) and an internal validation set (n=65) at an 8∶2 ratio, and the external test set (n=60) was derived from the other two independent centers. Two-dimensional (2D), 2.5D triplanar, three-dimensional (3D) deep learning features, radiomics features, and clinical features were extracted from pretreatment contrast-enhanced T1-weighted MRI. All types of features were mapped via a unified dense neural network (DNN) encoder, followed by a multitask prediction head to jointly output NACr classification and EFS survival estimation. Single-plane ablation and multitask ablation experiments were performed to verify the key design of the model, and the robustness of the model was evaluated in subgroups stratified by 7 clinical variables including age and pathological subtype. Results: For NACr prediction by the 2.5D multitask model, the area under the curve (AUC) was 0.818 (95%CI: 0.712-0.916) in the internal validation set and 0.770 (95%CI: 0.646-0.897) in the external test set, with negative predictive values of 92.6% and 85.7%, respectively. For EFS estimation, the concordance index (C-index) of the model was 0.779 (95%CI: 0.650-0.905) in the internal validation set and 0.732 (95%CI: 0.586-0.856) in the external test set, which was superior to the 2D model, 3D model, radiomics model, and clinical feature model. Ablation experiments confirmed that both triplanar integrated fusion and multitask joint learning were key factors for the model performance. Subgroup analysis showed no statistically significant difference in prediction performance among all clinicopathological subgroups (all P>0.05). The EFS risk score derived from the 2.5D model could effectively further stratify the survival prognosis in both NACr-sensitive and NACr-resistant subgroups (all P<0.05). Conclusion: The MRI-based deep learning multitask model can simultaneously and effectively predict chemotherapy response and EFS in children with head and neck RMS before treatment.