Background: MRI is widely used in the clinical evaluation of colorectal cancer because of its high resolution in soft tissue imaging, especially in the display of intestinal wall structure and invasion depth. MRI also can complete multi-angle, multi-parameter, and multi-sequence imaging. The prognosis of colorectal cancer patients with different depths of invasion varies significantly with different T-stage. The 5-year survival rate of patients with extramural invasion depth of less than 5mm is 85.4%, while the 5-year survival rate decreases to 54.1% when the depth of invasion is more than 5mm. According to the National Comprehensive Cancer Network (NCCN) rectal cancer Clinical Practice guidelines, patients with early-stage (T1-2N0) rectal cancer can undergo surgery directly. Locally advanced (T3-4) and T1-2 (T1-2N+) patients with regional lymph node metastasis need neoadjuvant chemoradiotherapy before surgery. Therefore, it is essential to determine the potential of different MRI modes to predict the T-stage and to efficiently distinguish T1-2 and T3-4 for patients to develop precise treatment plans.

Objective: The aim is to establish an efficient prediction system for colorectal cancer T-staging and identify each MRI mode's predictive potential.

Methods: A new multi-scale attention convolutional network based on Res2net and CBAM is proposed, named Res2net+CBAM. Res2net+CBAM extracts MRI features to predict the T-stage of colorectal cancer by fine-tuning the network and introducing the CBAM to focus model weight. Firstly, the multimodal MRI and the predictive performance of FCViT, Resnet, and Res2net for T staging of colorectal cancer were analyzed, revealing the correlation between the performance of FCViT and the data volume and the superiority of Res2net. Finally, the attention mechanism module CBAM is introduced, which can represent more multi-scale information and increase the receptive field.

Results: The study trained, tested, and validated T-stage in 405 colorectal cancer patients enrolled at local Hospital. 6 MRI sequences (MRI sequences included: OAx T2, DWI, OAx LAVA, OAx  T1, OSag LAVA and OSag T2) are included. Res2net+CBAM model improved the T-stage prediction performance of 6 MRI sequences to different degrees, and the improvement was more obvious in OAx T2, T1, and OSag T2 sequences, improved by 4.55%, 6.00%, and 4.18%, respectively. OAx LAVA and OSag T2 sequences predict T-stage better than other sequences. The testing set and validation set accuracy of OAx LAVA are 96.60% and 95.16%.The testing set and validation set accuracy of OSag T2 are 97.87% and 94.58%. In addition, a comparative analysis of  FCViT、Resnet, and Res2net reveals that the performance of the FCViT model was poor and related to the data volume.

Conclusions: In conclusion, Res2net+CBAM is a lightweight, interpretable, and efficient method for predicting colorectal T-stage for achieving precision treatment for patients with colorectal cancer by assisting clinicians in therapeutic decision-making. In addition, artificial intelligence models themselves have certain requirements for data volume.