MST-Net: a multi-scale spatial transformation network for automated T2/T3 staging of rectal cancer on preoperative T2-weighted MRI

OBJECTIVE: This study aimed to develop a deep learning model (MST-Net) for the segmentation-free automated and precise differentiation between T2 and T3 staging of rectal cancer based on preoperative T2-weighted magnetic resonance imaging (T2WI). MATERIALS AND METHODS: Preoperative T2WI images and postoperative pathologically confirmed T2/T3 staging labels from rectal cancer patients were retrospectively collected from two centers between January 2020 and January 2025. The core of MST-Net includes: a pyramid cross-stage feature fusion module to integrate shallow texture details and deep semantic information, mitigating the loss of spatial details in deep networks; and a lightweight spatial attention mechanism to adaptively focus on key radiomics features within the tumor region. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and accuracy (ACC). RESULTS: A total of 309 patients were included in the analysis, comprising a training set (n = 216) and an internal test set (n = 54) from Center 1, and an external independent validation set (n = 39) from Center 2. MST-Net achieved an AUC of 0.95 (ACC = 90%) on the internal test set, outperforming junior radiologists (ACC = 69%) and performing comparably to senior radiologists (ACC = 85%). On the external validation set, it achieved an AUC of 0.85 (ACC = 79%), also surpassing junior radiologists (ACC = 65%) and reaching the level of senior radiologists (ACC = 72%). CONCLUSION: MST-Net enables segmentation-free automated and high-accuracy prediction of T2/T3 staging in rectal cancer based on routine T2WI images. This model provides a reproducible auxiliary tool for clinical personalized treatment decision-making, potentially reducing the dependence of staging results on physician experience. KEY POINTS: Question Accurate differentiation between T2 and T3 rectal cancer on preoperative T2-weighted MRI remains clinically challenging but is essential for treatment planning. Findings MST-Net achieved high diagnostic accuracy and robust discrimination for T2 versus T3 staging in both internal and external validation cohorts. Critical Relevance This study demonstrates that a segmentation-free deep learning model using routine T2-weighted MRI can enable accurate preoperative T-staging of rectal cancer, supporting standardized and efficient radiological assessment and facilitating more consistent clinical decision-making in rectal cancer management.