Attention-Based Deep learning Network for automatic detection of Microsatellite Instability in endoscopy images

Abstract Background Microsatellite instability (MSI) is a clinically relevant biomarker in colorectal cancer (CRC). In this study, we employed a deep convolutional neural network (DCNN) model to differentiate between MSI and microsatellite-stable (MSS) status using endoscopic images. Methods A pathology-proven dataset was constructed from 986 colonoscopy images of 176 patients with CRC. The images were partitioned into training, validation, and test sets. Attention-based ResNet-50 and MobileNet-V2 DCNN architectures were trained and evaluated for their discriminatory performance. The area under the curve (AUC) was used as the primary performance metric, supplemented by accuracy, sensitivity, specificity, and precision. Results In the test cohort, the MobileNet-SE model achieved the best performance, with an AUC of 0.874 (95% CI: 0.852–0.896), a precision of 0.830 (95% CI: 0.787–0.863), a sensitivity of 0.714 (95% CI: 0.670–0.754), a specificity of 0.852 (95% CI: 0.815–0.882), and an accuracy of 0.782 (95% CI: 0.754–0.808). Conclusions These findings underscore the clinical utility of endoscopic images in MSI screening and establish a benchmark for an optimal MSI prediction model.