Computer-Aided Diagnosis for Lesion Treatability With White-Light Endoscopy

Surgical room in hospital with robotic technology equipment, machine arm surgeon in futuristic operation room. DaVinci robot. Minimal invasive surgical innovation, medical robot surgery with endoscopy
The aim of this study was to create and assess a new computer-aided diagnosis (CAD) system using only white-light endoscopy to determine whether to use endoscopic treatment for colorectal cancer lesions.

A computer-aided diagnosis (CAD) system could assist clinicians in making treatment decisions for colorectal lesions. These findings, from a retrospective diagnostic study, were published in Gastroenterology Endoscopy.

Investigators collected images (n=3442) extracted during white-light endoscopy that had not been magnified or stained from 1035 colorectal lesions (low-grade dysplasia [LGD], n=105; high-grade dysplasia [HGD], n=377; colorectal cancer [CRC] with submucosal invasion <1000 μm, n=107; CRC with submucosal invasion ≥1000 μm, n=146; advanced CRC, n=300). They collected the images at the Chiba University Hospital or Chiba Cancer Center between 2007 and 2018. Researchers trained the CAD device using a subset of 2751 images, and the test set consisted of the remaining 691 images. They compared the results from the test set images with the analysis from 2 trainee and 2 expert clinicians.

The CAD system was able to distinguish between treatable and untreatable lesions with an area under the receiving operating characteristic curve of 0.913 (sensitivity, 96.7%; specificity, 75%; positive predictive value, 90.2%; negative predictive value, 90.5%; accuracy, 90.3%).

Compared with the clinician diagnoses, the CAD outperformed the trainees for sensitivity (96.7% vs 92.1%; P <.01), specificity (75% vs 67.6%; P <.01), accuracy (90.3% vs 84.9%; P <.01), and time for diagnosis (0.009 vs 2 seconds; P <.01), respectively. The CAD was not significantly better at distinguishing between treatable and untreatable lesions compared with the expert opinions (sensitivity: P =.888; specificity: P =.337; accuracy: P =.394), but the CAD diagnosed images more quickly (0.009 vs 1.64 seconds; P <.01).

The investigators separated images by lesion type and observed that compared with trainees, the CAD was better at diagnosing HGD (95.9% vs 92.9%; P <.01), CRC with submucosal invasion <1000 μm (96.4% vs 80.7%; P <.01), and CRC with submucosal invasion ≥1000 μm (51.2% vs 31.5%; P <.01). The CAD was not superior at diagnosing LGD or advanced CRC than the trainees. The CAD performed similarly as the experts at diagnosing all types of lesions, with the exception of CRC with submucosal invasion ≥1000 μm, where the CAD outperformed the experts (51.2% vs 41.1%; P =.047).

A major limitation of this study was that it assessed whether lesions were able to be treated during endoscopy; however, the CAD was designed to be used for still images, indicating that at this stage, the system could not be implemented for real-time diagnosis during the endoscopy.

The study authors concluded that this computerized model had the ability to improve or aid in the diagnosis for treatment of lesions during a white-light endoscopy, especially for difficult lesions to diagnose, such as CRC with submucosal invasion ≥1000 μm.

Follow @Gastro_Advisor


Tokuaga M, Matsumura T, Nankinzan R, et al. A computer-aided diagnosis system using only white-light endoscopy for the prediction of invasion depth in colorectal cancer. Gastrointest Endosc. doi: 10.1016/j.gie.2020.07.053