A computer-aided diagnosis (CAD) may aid in the detection and characterization of colorectal lesions, according to results from a machine-learning study published in Gastrointestinal Endoscopy.
Researchers from the University of Adelaide developed CAD software using a machine-learning approach. Their computational model was validated using a dataset of 1235 polyps (narrow-band images [NBIs]) retrospectively collected from a hospital in Australia between 2010 and 2016. The images included 103 hyperplastic polyps, 429 low-grade tubular adenomas, 293 low-grade sessile serrated adenomas, 295 high-grade colorectal lesions, and 115 invasive colorectal cancers. For validation, the researchers tested their model on a set of 20 narrow-band polyp images and 49 blue-laser images (BLIs) collected in Japan.
The CAD software performed well on the test set, with an area under the curve (AUC) of 94.3% (95% CI, 93.3%-96.2%) when testing 4 images per lesion type. The validation set had similarly high AUC scores of 84.5% (95% CI, 78.3%-88.6%) for the NBIs and 90.3% (95% CI, 87.9%-92.8%) for the BLIs, when testing 1 image per lesion type. The investigators did not report any significant difference of the CAD model performance for image or lesion type.
A major limitation of this study is the small validation sample size. The validation set had only a few images of each lesion type (e.g., 5 hyperplastic polyps for NBIs and 9 for BLIs), resulting in the validation stage of the study testing only 1 image at a time. Furthermore, the researchers were unable to include all colonoscopy imaging technologies in the study. For these reasons it remains unclear whether this CAD may be useful for wide use.
The authors of the study concluded that their CAD model had comparable AUC with experts and could be used for the 2 imaging types tested (NBIs and BLIs). They asserted that their promising results might, in the future, improve detection and classification of colorectal lesions.
Cheng Tao Pu LZ, Maicas G, Tian Y, et al. Computer-aided diagnosis for characterisation of colorectal lesions: a comprehensive software including serrated lesions [published online March 4, 2020]. Gastrointest. Endosc. doi:s10.1016/j.gie.2020.02.042