A context-aggregated attention model for predicting KRAS mutation status in colorectal cancer

Purpose: Accurate identification of kratom sarcoma(KRAS) gene mutation status is of great significance for prognosis and treatment response in colorectal cancer. However, MRI images of colorectal cancer suffer from complex background, variable shape of the same type of lesion, and current network models do not conform to end-to-end learning, so accurate identification of KRAS gene mutation status is still challenging. Methods: In this research, we propose a Context-Aggregated Attention Model(CAAM) for predicting KRAS mutation status in colorectal cancer. The model's general framework comprises a segmentation framework and a classification framework, where segmentation aids in the classification process. Initially, the pre-processed MRI images are fed into the modified TransUNet to generate segmentation feature maps. These segmented feature maps are then combined with the preprocessed image to create a bootstrap feature map. The bootstrap feature map guides Mask-Guided Module(MGM) and the Contextual Aggregation Convolution module(CAC) to derive the final classification results. To train the network, we employ a combined loss function that incorporates both segmentation and classification losses. Results: : We trained the network using a ten-fold cross-validation method and evaluated our model using the T2-weighted MRI dataset. The experiments demonstrated that our model achieved an accuracy of 87.56% on the test dataset, outperforming the results of the comparative experiments. Conclusion: The results indicate that our proposed context-aggregated attention model can provide new perspectives and approaches for the auxiliary treatment of colorectal cancer.