MultiCubeNet: multitask deep learning for molecular subtyping and prognostic prediction in gliomas

Gliomas, the most prevalent type of primary brain tumors, require precise molecular characterization for effective diagnosis and treatment. Despite advancements in radiomics, simultaneous prediction of key molecular markers, such as isocitrate dehydrogenase (IDH) mutation, 1p/19q co-deletion, and telomerase reverse transcriptase (TERT) promoter mutation, along with prognosis, remains challenging. We aimed to develop and validate a deep learning (DL) model capable of simultaneously predicting key genetic molecular markers and prognosis in gliomas.

We conducted a retrospective analysis of 457 adult-type diffuse gliomas (193 training cohort; 162 and 102 cases in SZS and TCGA validation cohorts, respectively). We developed MultiCubeNet, a multisequence, multiscale, multitask deep learning framework designed to predict IDH mutation, 1p/19q co-deletion, TERT promoter mutation, and prognosis. Model performance was benchmarked against conventional radiomics pipelines and neuroradiologist annotations. Classification accuracy was evaluated by the area under the receiver operating characteristic curve (AUC), with prognostic performance quantified using Harrell's concordance index (C-index).

The median age of the patients was 49 years, and 266 were men (58.2%). The model demonstrated high efficiency in the training set, achieving AUCs of 0.966 for IDH mutation, 0.961 for 1p/19q co-deletion, and 0.851 for TERT promoter mutation. In the external test set (SZS), the model maintained strong performance with AUCs of 0.877, 0.730, and 0.705 for IDH mutation, 1p/19q co-deletion, and TERT promoter mutation, respectively. The performance in TCGA cohort was less optimal, with AUCs below 0.8. The framework consistently matched or exceeded both radiomics pipelines and neuroradiologists in molecular marker identification. Survival analysis revealed significant prognostic stratification across all cohorts (C-index: 0.706-0.866).

MultiCubeNet, a multitask deep learning model leveraging multisequence and multiscale MRI, demonstrated strong performance in predicting key molecular markers and prognosis in gliomas, thereby supporting personalized treatment approaches.