AI could help detect colorectal cancer on a routine abdominopelvic computed tomography (APCT) scan without bowel preparation, a finding that could help detect more cancer cases at an earlier stage, according to a presentation on Tuesday.
Kim
Colorectal cancer is an easily overlooked intraabdominal malignancy on APCT when unsuspected.
"I started to think that if I could get any help from AI, the very first thing I would like to receive was the automatic detection of colorectal cancer," said Seung-Seob Kim, MD, clinical assistant professor, Department of Radiology and Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine in South Korea. "As I searched for previous studies regarding the automatic detection of colorectal cancer, however, I was surprised that the majority of previous works on the computer-aided detection of colorectal cancer focused only on either optical colonoscopy videos or CT colonography images."
Routine APCT Could Use AI To Detect Cancer Sooner
The team developed an AI algorithm to automatically detect colorectal cancer on routine abdominopelvic computed tomography (APCT) scans without bowel preparation. The AI model was trained using APCTs from 2,662 patients diagnosed with colorectal cancer between 2010 and 2014. The algorithm was then validated internally (841 patients) and externally (442 patients) with datasets of consecutive patients with or without colorectal cancer who underwent APCT and optical colonoscopy within a two-month interval at two independent tertiary hospitals between January and June 2018.
Results showed nearly zero false positive detections of cancer, showing the potential feasibility of AI-based algorithm for detecting colorectal cancer in APCT scanned without bowel preparation, Dr. Kim said. Among the 841 patients in the internal validation set, 92 patients were diagnosed with colorectal cancer by both the colonoscopy and APCT.
Part of CT images from a 72-year-old female patient who was histologically diagnosed with rectal adenocarcinoma are shown. Among eleven slices with labeled ground truth boxes, AI predicted seven bounding boxes contiguously with inner overlapping, which were therefore regarded as one single lesion. The sum of predicted probabilities of the seven boxes was 4.7464. The average DSC was 0.7947, which was calculated by dividing the sum of DSCs by seven, the total number of slices of the AI-predicted lesion. As average DSC was greater than 0.3, this lesion was regarded as true positivity.
Radiologists Often Aren't Looking for Cancer on APCT Unless Indicated
The biggest difference between CT colonography and routine APCT lies in the purpose of the exam. CT colonography is performed primarily to detect colorectal polyp or cancer, whereas routine APCT is indicated for various clinical purposes not limited to colorectal cancer detection. That's why routine APCTs generally outnumber CT colonography exams.
The researchers hypothesized that if colorectal cancer can be automatically detected on routine unprepared APCT images, the clinical usefulness may be much greater than previously proposed.
"For example, let's imagine a patient coming to the emergency department due to acute symptoms such as high fever or unspecified abdominal pain. A clinician decides to perform a routine APCT, and a radiologist may want to try his or her best to find causes of fever or pain. In such a situation, colorectal cancer can be easily missed," Dr. Kim said. "If we had an AI algorithm capable of detecting unsuspected, colorectal cancer regardless of the reason the CT was originally performed, patients could be diagnosed with colorectal cancer earlier."
Dr. Kim said the AI model still needs to be prospectively validated and that the specificity of the AI model needs some improvement, but it could still be beneficial for radiologists in the future as they look to improve patient care and diagnose cancer sooner.
Access the presentation, "Development of The Artificial Intelligence-based Algorithm for Detecting Colorectal Cancer Using An Unprepared Abdominopelvic Computed Tomography" (T6-SSGI10-3) on demand a Meeting.RSNA.org.
