P0924 - Deep Learning and Capsule Endoscopy: Automatic Identification and Differentiation of Small Bowel Lesions With Distinct Hemorrhagic Potential Using a Convolutional Neural Network

Sunday, October 24, 2021

3:30 PM – 7:00 PM PT

Location: Shoreline Exhibit Hall

Has Audio

Award: Presidential Poster Award

Presenting Author(s)

Miguel Mascarenhas, MD

Centro Hospitalar de São João
Porto, Porto, Portugal

Award: Presidential Poster Award

Miguel Mascarenhas, MD¹, Joao Afonso, MD¹, Tiago Ribeiro, MD¹, João Ferreira, PhD², Hélder Cardoso, MD¹, Patrícia Andrade, MD¹, Marco Parente, PhD², Renato Natal, PhD², Guilherme Macedo, MD, PhD, FACG¹
¹Centro Hospitalar de São João, Porto, Porto, Portugal; ²University of Porto, Porto, Porto, Portugal

Introduction: Capsule endoscopy (CE) has revolutionized the investigation of patients with suspected small bowel disease. Obscure gastrointestinal bleeding is the most common indication for CE and, most frequently, it originates from the small bowel. Capsule endoscopy frequently identifies a wide range of lesions with different bleeding potential in these patients. However, reading CE exams is a time-consuming and monotonous task which is prone to errors. Convolutional neural networks (CNNs) are highly efficient artificial intelligence tools for image analysis. We aimed to develop a CNN-based model for identification and differentiation of multiple small bowel lesions with distinct hemorrhagic potential using CE images.

Methods: Our group developed, trained, and validated a denary CNN based on CE images. Each frame was labeled according to the type of lesion (lymphangiectasia, xanthomas, ulcers, erosions, vascular lesions, protruding lesions, and blood) by three experts in CE. The hemorrhagic potential was assessed by Saurin’s classification: P0 – lesions without bleeding potential; P1 – lesions with uncertain bleeding potential; P2 – lesions with high bleeding potential; P3 – luminal blood. A training dataset was used for development of the model (80% of the full image dataset). The CNN’s prediction was compared to the experts’ labeling (Figure 1a). The performance of the network was evaluated using an independent dataset (20% of the full image dataset). The performance of the CNN was measured by the area under the receiving operating characteristic curve (AUROC), sensitivity, specificity, positive and negative predictive values (PPV and NPV, respectively).

Results: A total of 53555 CE images were included from 1229 patients (table 1). The model had an overall accuracy of 99% (Figure 1b) a sensitivity of 88%, a specificity of 99%, a PPV of 87% and NPV of 99% for detection of multiple small bowel abnormalities and respective classification of bleeding potential. The AUROCs for the detection of the different types of lesions is shown in Figure 1c.

Discussion: Our group successfully developed and tested a CNN-based model for automatic detection and differentiation of multiple types of small bowel lesions and classification of the respective bleeding potential. This system may improve the diagnostic yield of CE for these lesions and overall CE efficiency. The novelty of this study resides on the CNN ability to not only detect but also classify multiple pleomorphic small bowel lesions.

Figure: Figure 1: 1a - Output obtained from the application of the convolutional neural network. A blue bar represents a correct prediction. A red bar represents an incorrect prediction 1b - Evolution of the accuracy of the convolutional neural network during training and validation phases, as the training and validation datasets were repeatedly inputted in the neural network. 1c - ROC analyses for detection of different types of lesions. CNN – convolutional neural network; N – normal mucosa; P0L – lymphangiectasia; P0X – xanthomas; P1E – mucosal erosions; P1U – P1 ulcers; P2U – P2 ulcers; P1RS – P1 red spots; P2V – P2 vascular lesions; P1P – P1 protruding lesions; P2P – P2 protruding lesions. The prefixes P0, P1 and P2 refer to the hemorrhagic potential according to Saurin’s classification.

		Experts’ classification
		N	P0L	P0X	P1E	P1U	P2U	P1RS	P2V	P1P	P2P	Blood
CNN’s prediction	N	3313	65	40	43	34	7	52	9	14	2	23
	P0L	52	932	3	16	4	0	2	3	5	2	1
	P0X	25	1	319	0	1	0	0	1	0	1	0
	P1E	38	8	0	307	27	4	4	0	1	0	2
	P1U	13	0	1	38	362	9	2	0	0	0	0
	P2U	7	0	0	5	11	466	3	0	0	0	1
	P1RS	56	1	2	9	4	2	303	8	0	0	2
	P2V	27	4	1	1	0	4	15	255	1	1	8
	P1P	56	43	8	3	3	2	1	4	338	5	2
	P2P	4	0	0	0	0	0	0	1	3	199	1
	Blood	11	1	0	0	1	1	0	0	0	1	3040
	Total	3602	1055	374	422	447	495	382	281	362	211	3080

Table: Table 1 – Confusion matrix of the automatic detection vs. experts’ classification

Disclosures:

Miguel Mascarenhas indicated no relevant financial relationships.

Joao Afonso indicated no relevant financial relationships.

Tiago Ribeiro indicated no relevant financial relationships.

João Ferreira indicated no relevant financial relationships.

Hélder Cardoso indicated no relevant financial relationships.

Patrícia Andrade indicated no relevant financial relationships.

Marco Parente indicated no relevant financial relationships.

Renato Natal indicated no relevant financial relationships.

Guilherme Macedo indicated no relevant financial relationships.

Miguel Mascarenhas, MD¹, Joao Afonso, MD¹, Tiago Ribeiro, MD¹, João Ferreira, PhD², Hélder Cardoso, MD¹, Patrícia Andrade, MD¹, Marco Parente, PhD², Renato Natal, PhD², Guilherme Macedo, MD, PhD, FACG¹. P0924 - Deep Learning and Capsule Endoscopy: Automatic Identification and Differentiation of Small Bowel Lesions With Distinct Hemorrhagic Potential Using a Convolutional Neural Network, ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.