CT Study Says Deep Learning Model Could Help Differentiate Between Acute Diverticulitis and Colon Carcinoma

Noting that overlapping imaging features on contrast-enhanced computed tomography (CT) can make it challenging to differentiate between acute diverticulitis and colon cancer, researchers say an emerging deep learning model may provide enhanced sensitivity and specificity for these conditions.

In a retrospective study recently published in JAMA Network Open, researchers developed and tested a three-dimensional (3D) convolutional neural network (CNN) for 585 patients (mean age of 63.2) who underwent surgery for colon cancer or acute diverticulitis between July 1, 2005 and October 1, 2020, had venous phase CT imaging within 60 days prior to surgery and had segmental wall thickening in the colon that was independent of disease stage.

In comparison to mean sensitivity and specificity rates of 77.6 percent and 81.6 percent, respectively, for radiologist readers, the study authors noted an 83.3 percent sensitivity rate and an 86.6 percent specificity rate for the 3D CNN model. The combination of the deep learning model and radiologist assessment resulted in an eight percent increase in sensitivity (85.6 percent) and a 9.7 percent increase in specificity (91.3 percent) over radiologist assessments, according to the study findings.

The study authors also noted the reduction of false-negative rates with the 3D CNN model. According to the study, the overall false-negative rate for radiology readers in the study decreased from 22 percent to 14.3 percent with adjunctive use of the 3D CNN algorithm. Specifically, the 3D CNN model led to a 4 percent reduction in false-negative rates (from 14 percent to 10 percent) for board-certified radiologists and a 9.9 percent reduction (from 26 percent to 16.1 percent) for radiology residents.

The researchers said the reduction in false-negative findings has “major clinical implications” for patients with colon cancer or acute diverticulitis.

“In the perforated stage, both entities require emergency surgery; however, the surgical strategies differ. Whereas (colon cancer) requires oncologic resection of the diseased bowel and the entire lymph node basin, a limited resection of the diseased bowel may suffice in cases of (acute diverticulitis). A high level of certainty in surgical planning improves patient stratification and thus limits postoperative complications and potentially decreases mortality rates,” wrote study co-author Rickmer Braren, M.D., who is affiliated with the Institute of Diagnostic and Interventional Radiology at the School of Medicine at the Technical University of Munich in Germany, and colleagues.

(Editor’s note: For related content, see “Diagnosing Pancreatic Lesions on Abdominal CT: Study Says Deep Learning System is Comparable to Radiologist Assessment” and “Could a New Deep Learning Tool Enhance CT Detection of Pancreatic Cancer?”)

In cases of early-stage colon cancer and acute diverticulitis, the study authors cautioned that subtle CT findings, such as adjacent fat stranding and focal bowel wall thickening, can be mistaken for peristaltic activity or be obscured by bowel filling. Braren and colleagues also noted that secondary changes like mesenteric stranding and abscess formation can be dominating features on CT in cases involving advanced colon cancer or complicated acute diverticulitis.

In regard to study limitations, the study authors acknowledged that broader application of the study results may be limited due to the AI model being trained and tested on a single institutional data set. They also noted that imaging features, such as fat stranding, may have been masked by adversarial noise (at a variance threshold of .01) that impacted the performance of the AI model. Noting that the study focused on the most frequent benign and malignant entities for bowel wall thickening, the study authors maintained that future studies should assess a broader array of malignant and benign entities and incorporate multiparametric data integration in order to evaluate and possibly improve the capabilities of the AI model.