Multi-sequence pituitary tumor MRI–based deep learning model for predicting aggressive acromegaly

Abstract Objectives Growth hormone-secreting pituitary adenomas (GHPAs) can exhibit highly aggressive oncological behaviors, characterized by local invasion, high recurrence rates, and resistance to standard multimodal therapies including surgery and radiotherapy. While these refractory tumors drive the systemic syndrome of acromegaly, early identification of their aggressive phenotype remains a major clinical challenge. To overcome the limitations of conventional morphological markers, this study aimed to develop a deep learning model using multi-sequence MRI to predict aggressive GHPAs preoperatively by capturing complex tumor image features. Methods In this single-center retrospective study, 102 patients with GHPAs were included, of which 38 exhibited aggressive, treatment-refractory behavior. Model inputs comprised multi-sequence pituitary MRIs: coronal contrast-enhanced T1-weighted (CE-T1), sagittal CE-T1, and coronal T2-weighted imaging. We implemented a hierarchical multiple instance learning framework to construct a multi-sequence residual fusion network (PTNet). This architecture featured shared-weight feature extraction, slice-level attention aggregation, and sequence-level residual fusion to automatically learn high-dimensional oncological imaging phenotypes from patient-level labels. Model performance was evaluated via 5-fold cross-validation, alongside decision curve and calibration analyses. Results During 5-fold cross-validation, PTNet achieved an area under the curve (AUC) of 0.9126, an accuracy of 0.8922, and a sensitivity of 0.8684 in identifying aggressive tumors, outperforming both single-sequence models and fusion models based on simple concatenation. Notably, patients in the aggressive group also exhibited significantly higher oncological proliferation markers, such as the Ki-67 index. Decision curve analysis demonstrated a favorable net clinical benefit across relevant risk thresholds, while calibration analysis indicated good agreement between predicted and observed probabilities. Conclusions The deep learning-based integration of multi-sequence MRI effectively captures latent imaging phenotypes associated with aggressive GHPAs. Although requiring external validation, this approach shows strong potential as a preoperative neuro-oncological tool to identify patients at high risk of refractoriness, thereby facilitating individualized multimodal interventions such as targeted surgical and adjuvant radiotherapy planning.