The emerging role of AI in gastroesophageal cancer

The role of artificial intelligence (AI) in the detection of upper gastrointestinal (GI) neoplasia has been emerging in recent years. Dr Wai Leung (University of Hong Kong, China) discussed the latest evidence on AI and the diagnosis, characterisation, risk prediction, and the examination of upper GI neoplasia [1].

A retrospective meta-analysis demonstrated that AI was accurate in detecting gastric neoplastic lesions, with an area under the receiver-operating characteristic curve (AUC) of 0.96, a sensitivity of 0.92, and a specificity of 0.88. For the detection of Barret’s oesophagus neoplasia, the corresponding results were 0.96, 0.88, and 0.90. For squamous oesophagus neoplasia the data showed an AUC of 0.92, a sensitivity of 0.84, and a specificity of 0.90 [2]. In addition, a prospective study including 1,198 patients showed a similar accuracy of endoscopic detection of gastric lesions and neoplasms by AI (0.72) and by expert (0.68). However, the AI model displayed a higher sensitivity (100% vs 85.5%; P=0.003) and negative predictive value (100% vs 86.4%; P=0.002) [3].

Furthermore, AI has been investigated for the purpose of quality control during endoscopy. A randomised controlled trial including 1,050 patients demonstrated that the use of AI during endoscopy significantly reduced the number of blind spots (mean 5.38 vs 9.82; P<0.001) [4]. Also, a crossover study comparing AI-assisted examination and routine examination of upper GI lesions showed that the miss rate per lesion was lower in the AI-first group (6.1%) than in the routine-first group (27.3%; P=0.015), resulting in a lower biopsy rate among patients who underwent AI-assisted examination first [5].

Another focus of AI research is endoscopy training. A study showed that feedback from AI improved the endoscopic results of endoscopists in training compared with their endoscopic decisions before AI feedback: negative predictive value (74.7% vs 82.5%; P=0.049), accuracy (69.3% vs 74.7%; P=0.003), and AUC (0.69 vs 0.75; P=0.02) [6].

A study by Ali et al. showed that AI could also improve the 3D quantification of Barrett’s oesophagus, which is helpful for further interventions and monitoring of these patients [7]. Also, the detection of oesophageal adenocarcinoma, a labour-intensive process, could be simplified using AI and thus reduce the manpower required [8]. Finally, AI was able to predict the risk of Barrett’s oesophagus (AUC 0.86) or gastric cancer development (AUC 0.90) in 2 recent studies [9,10].

Dr Leung concluded that these positive results need to be confirmed in future research, establishing the clinical applicability and cost-effectiveness of AI in the detection of upper GI neoplasia.

1. Leung WK, et al. Artificial Intelligence-Assisted Detection of Upper Gastrointestinal Neoplasia. Presentation 1, Breakout Session: Understanding Disparities and Expanding Access Through Diagnostic Technology and Treatment in Gastroesophageal Cancers, ASCO GI 2022, 20–22 January.
2. Lui TKL, et al. Gastrointest Endosc. 2020;92(4):821–830.
3. Wu L, et al. Gastrointest Endosc. 2022;95(2):269–280.e6.
4. Wu L, et al. Endoscopy. 2021;53:1199–1207.
5. Wu L, et al. Lancet Gastroenterol Hepatol. 2021;6(9):700–708.
6. Lui TKL, et al. Endosc Int Open. 2020;8(2):E139–E146.
7. Ali S, et al. Gastroenterology. 2021;161:865–878.e8.
8. Gehrung M, et al. Nat Med. 2021;27(5):833–841.
9. Rosenfeld A, et al. Lancet Digit Health. 2020;2(1):E37–48.
10. Leung WK, et al. Aliment Pharmacol Ther. 2021;53:864–872.

 

Copyright ©2022 Medicom Medical Publishers



Posted on 21 March, 202221 March, 2022