AN ARTIFICIAL INTELLIGENCE-BASED SYSTEM FOR AUTOMATICALLY MEASURING THE SIZE OF ENDOSCOPIC GASTROINTESTINAL LESIONS

Aims In the process of gastrointestinal endoscopy, it is necessary to measure the size of lesion, which is one of the bases for the risk classification and treatment of various diseases. We aimed at developing an artificial intelligence-based system for measuring the size of gastrointestinal lesions in real-time. We intended to assess the accuracy of the system on a 3D colon model and endoscopic patients.

Methods The system was trained by deep convolutional neural networks, which integrateed depth measurement and size assessment into a model. The system was developed using 5000 images. These images photographed a series of polypoid objects (2mm to 15mm) in vitro with object distance varied between 5mm and 100mm. The derived system was prospectively tested on a on a 3D-print colon model with polypoid objects and on 200 endoscopic videos from 170 patients.

Results The system demonstrated a 91.15% accuracy with a mean absolute error of 2.60mm on 3D print colon compared with ground truth. On patient data, the system achieved similar metrics across different size of lesions with pathology results as gold standard. The assessments were more stable and accurate than endoscopists.

Fig. 1

Conclusions The proposed system could automatically figure out depth and size of lesion. The novel methodology will quantify lesion assessment for all kinds of diseases, expanding the application of artificial intelligence on medicine.

Publication History

Article published online:
14 April 2022