Artificial intelligence may help in predicting the need for additional surgery after endoscopic resection of T1 colorectal cancer

Abstract

Background and study aims Decisions concerning additional surgery after endoscopic resection of T1 colorectal cancer (CRC) are difficult because preoperative prediction of lymph node metastasis (LNM) is problematic. We investigated whether artificial intelligence can predict LNM presence, thus minimizing the need for additional surgery. Patients and methods Data on 690 consecutive patients with T1 CRCs that were surgically resected in 2001 – 2016 were retrospectively analyzed. We divided patients into two groups according to date: data from 590 patients were used for machine learning for the artificial intelligence model, and the remaining 100 patients were included for model validation. The artificial intelligence model analyzed 45 clinicopathological factors and then predicted positivity or negativity for LNM. Operative specimens were used as the gold standard for the presence of LNM. The artificial intelligence model was validated by calculating the sensitivity, specificity, and accuracy for predicting LNM, and comparing these data with those of the American, European, and Japanese guidelines. Results Sensitivity was 100 % (95 % confidence interval [CI] 72 % to 100 %) in all models. Specificity of the artificial intelligence model and the American, European, and Japanese guidelines was 66 % (95 %CI 56 % to 76 %), 44 % (95 %CI 34 % to 55 %), 0 % (95 %CI 0 % to 3 %), and 0 % (95 %CI 0 % to 3 %), respectively; and accuracy was 69 % (95 %CI 59 % to 78 %), 49 % (95 %CI 39 % to 59 %), 9 % (95 %CI 4 % to 16 %), and 9 % (95 %CI 4 % – 16 %), respectively. The rates of unnecessary additional surgery attributable to misdiagnosing LNM-negative patients as having LNM were: 77 % (95 %CI 62 % to 89 %) for the artificial intelligence model, and 85 % (95 %CI 73 % to 93 %; P < 0.001), 91 % (95 %CI 84 % to 96 %; P < 0.001), and 91 % (95 %CI 84 % to 96 %; P < 0.001) for the American, European, and Japanese guidelines, respectively. Conclusions Compared with current guidelines, artificial intelligence significantly reduced unnecessary additional surgery after endoscopic resection of T1 CRC without missing LNM positivity.

Artificial Intelligence in Medicine: Today and Tomorrow

“…Gastroenterologists made use of convolutional neural networks among other deep learning models in order to process images from endoscopy and ultrasound (32) and detect abnormal structures such as colonic polyps (33). Artificial neural networks have also been used to diagnose gastroesophageal reflux disease (34) and atrophic gastritis (35), as well as to predict outcomes in gastrointestinal bleeding (36), survival of esophageal cancer (37), inflammatory bowel disease (38), and metastasis in colorectal cancer (39) and esophageal squamous cell carcinoma (40).…”

Clinicopathological features of T1 colorectal carcinomas with skip lymphovascular invasion

“…With recent advances in endoscopic treatments, such as ESD, many T1 CRCs can be resected endoscopically with negative margin. Although endoscopic treatment is an attractive option for the patients, T1 CRCs show approximately 10% of LNM, which requires surgical colectomy with lymph node dissection to cure the cancer (13)(14)(15)(16)(17)(18)(19)(20). Therefore, it is important to reveal the risk factors for LNM in the patients with T1 CRCs, and we should decide whether additional surgery after endoscopic resection is needed according to the presence of risk factors for LNM.…”

Are Morphologic and Histologic Risk Factors for T1 Colorectal Cancer Friends or Enemies?

“…Fourth, prospective studies still are needed because the role of morphology has not been confirmed in other settings. 11 In conclusion, we must admit that the convergence between the protective effect of the pedunculated morphology, on one side, and the much lower prevalence of histologic risk factors in pedunculated T1 CRCs, on the other, somewhat simplifies the incorporation of morphology in the clinical management of these lesions, further supporting the use of different management approaches between pedunculated and nonpedunculated T1 CRCs.…”