A multimodal MRI-based deep learning model for non-invasive diagnosis of pineal region germinoma

To develop a deep learning model for the preoperative, non-invasive identification of germinomas in the pineal region.

This retrospective study included 114 patients with pathologically confirmed pineal region tumors. The cohort was randomly divided into a training set (n = 91) and a test set (n = 23). The training set was further partitioned into three folds for cross-validation. A convolutional neural network (CNN) enhanced with contrastive learning was used to extract discriminative features from individual MRI sequences and demographic data. A mixed-attention mechanism was then employed to fuse these features into multimodal representations, aiming to improve identification performance.

In this retrospective cohort, the mean age was 9.63 ± 4.99 years, with a male predominance (85.09%). Germinomas accounted for 51.75% of cases. Significant differences were observed between the germinoma and non-germinoma groups in age (p = 0.001) and sex (p = 0.046). Among traditional single-modality models, the T1CE model showed the best performance (test set accuracy [ACC] 0.667, area under the curve [AUC] 0.803), which improved with contrastive learning (ACC 0.797, AUC 0.848). The hybrid attention-based multimodal fusion model achieved superior discriminative performance (test set ACC 0.855, AUC 0.919), significantly outperforming all single-modality models.

By integrating multimodal MRI features, the model demonstrates promising performance for the preoperative identification of germinomas. This non-invasive approach may help guide clinical decision-making in patients with pineal region tumors.