Predicting Lymph Node Metastasis of Colorectal Cancer in CT Scans Using Attention-Based Multiple Instance Learning.

Accurately predicting lymph node metastasis (LNM) in colorectal cancer (CRC) contributes to develop appropriate treatment plans and evaluate postoperative prognosis. In clinical practice, preoperative diagnosis of LNM in CRC mostly relies on computed tomography (CT). Nevertheless, CT-based diagnosis of metastatic lymph nodes has a substantial misdiagnosis rate and lacks consistency among clinicians. Additionally, existing deep learning models using CT scans to predict LNM in CRC have not yielded satisfactory results due to insufficient attention given to lymph nodes. To address these issues, we propose an attention-based multiple instance learning (MIL) framework for this challenging task. Our framework incorporates a global-local cross-attention (GLCA) feature fusion module that combines global information from CT slices with local information from lymph nodes. Furthermore, an attention-based pooling approach is utilized to extract critical features and generate a bag representation. Lastly, a nested-neural memory ordinary differential equations (N-nmODE) feature enhancement module significantly augments the expressive capacity of the bag representation. A series of empirical studies show that our method achieves an overall accuracy of 74.7% and AUC of 76.8%, which represents a marked improvement over the physicians specialising in CRC and outperforms existing state-of-the-art methods. Our source code is available at https://github.com/SCU-MI/CAN-MIL.