A two-stage deep learning prediction system for colon cancer microsatellite instability status using CT images.

BACKGROUND

This study seeks to build a two-stage deep learning approach for identifying the microsatellite instability (MSI) status of colon cancer based on computed tomography (CT) scans without the requirement for manual segmentation.

METHODS

This study included 108 enhanced CT scans of colon cancer, including 68 cases of ascending colon, 14 cases of transverse colon, 18 cases of descending colon, and 8 cases of sigmoid colon; there were 56 cases of MSI-H and 52 cases of microsatellite stability (MSS). In the first stage, the segmentation model MSI-SAM was trained to accurately segment the lesion locations in the CT scans. In the second stage, the mask acquired from the MSI-SAM segmentation was multiplied by the original CT image (CT_Origin) bitwise, and the result was merged with the mask obtained from the MSI-SAM segmentation (Segment) to obtain CT_ROI. Both CT_ROI and CT_Origin were then diagnosed using the colon cancer MSI status diagnosis model.

RESULTS

The performance of the suggested CT segmentation model MSI-SAM in the ascending colon, transverse colon, descending colon, and sigmoid colon areas (DSC: IoU) was (0.886:0.798), (0.878:0.783), (0.923:0.857), and (0.854:0.747), respectively. The AUC of the MSI status diagnostic model for patients with colon cancer was 0.935 (95% CI 0.892-0.947), the ACC was 0.913, the sensitivity was 1.000, and the specificity was 0.846.

CONCLUSIONS

The segmentation masks created by the trained deep learning segmentation model achieved a level comparable to that of expert radiologists, and the deep learning diagnostic model played an essential role in supporting doctors in diagnosis.