Artificial intelligence in detection of small bowel lesions and their bleeding risk: A new step forward

Abstract

The present letter to the editor is related to the study with the title “Automatic detection of small bowel (SB) lesions with different bleeding risk based on deep learning models”. Capsule endoscopy (CE) is the main tool to assess SB diseases but it is a time-consuming procedure with a significant error rate. The development of artificial intelligence (AI) in CE could simplify physicians’ tasks. The novel deep learning model by Zhang et al seems to be able to identify various SB lesions and their bleeding risk, and it could pave the way to next perspective studies to better enhance the diagnostic support of AI in the detection of different types of SB lesions in clinical practice.

Key Words: Capsule endoscopy, Small bowel, Artificial intelligence, Bleeding risk, Vascular lesions

Core Tip: The development of artificial intelligence (AI) in capsule endoscopy could simplify physicians tasks by reducing the number of frames that need to be analyzed by the physician. The novel model proposed in the retrospective study by Zhang et al. could be able to identify various small bowel lesions and their bleeding risk. More prospective studies are needed to validate this promising AI-model.

TO THE EDITOR

We were intrigued by the recently published paper by Zhang et al[1] evaluating the accuracy of a new classification and detection deep learning model to identify different types of small bowel (SB) lesions and their bleeding risk. The proposed model showed high diagnostic sensitivity and accuracy, and it resulted in a significant reduction of physician’s reading time.

Capsule endoscopy (CE) is the main tool to assess SB diseases[2] and, since its introduction, the capsule readers spend on average from 30 to 120 min to review and interpret full-length CE videos containing thousands of frames, thus leading to a gradual reduction of reader attention. The development of artificial intelligence (AI) and its application in SB diseases detection, represents a hot topic, since it could simplify endoscopic examination by reducing the number of frames that need physicians review.

The deep learning systems and convolutional neural networks in AI-empowered CE increase its diagnostic yield, allowing the automatic detection and characterization of pathological lesions, with shorter reading time while maintaining high diagnostic accuracy. In the literature the AI-solution has showed a 98% sensitivity and 99% specificity in the identification of occult/obscure gastrointestinal bleeding sources, and an accuracy of 99.9%[3,4]. It should be emphasized that the existing studies are not able to evaluate the bleeding risk of SB lesions, which in clinical practice are fundamental given that decisions about further treatments mostly rely on this factor.

The authors’ proposed model in this retrospective study comprises a two-stage method which explores the image classification and object detection (Improved ResNet50 + YOLO-V5), and can simultaneously detect many SB lesions, providing an accurate localization, and assess the lesions’ bleeding risk. Training data and testing data were constructed respectively on 74574 and 37287 images, subsequently model-assisted readings were compared with physician readings. The diagnosis of at least three specialists was considered as the gold standard in cases where the final diagnosis was unclear or different lesions were identified. The diagnostic performance of this model showed greater sensitivity (99.17%) and accuracy (98.6%) and resulted in significant time-savings compared with standard reading time (48msvs0.4s/image).

Overtime, many models were developed to automatically identify different pathological SB lesions (for example ulcers, polypoid lesions, erosions), however the research on AI in CE is being challenged by their variability in depth, size, forms, and origin. Compared to the existing models[5,6], Zhang et al^[ showed a better performance in identification of ulcers, vascular lesions and bleeding.

Last but not least, this study’s remarkable originality was its excellent sensitivity and accuracy in supporting physician reading. The physician’s reading sensitivity was just 52.38%, which meant that almost half of the lesions would have been lost, particularly with SB vascular lesions; this result compared to the 99% sensitivity of the model, bears a huge significance for physicians who are responsible for improving bleeding diagnosis and identifying patients at high risk of bleeding.

We think that this experimental two-stage model could guide next studies (evaluating not only images in a retrospective approach) to better enhance the diagnostic support of AI in the detection of more various SB lesions in clinical practice, especially for young and inexperienced gastroenterologists. Before physicians can fully rely on such solutions and cut down on reading time while preserving (and potentially even increasing) detection accuracy, prospective controlled studies are still mandatory.