Artificial intelligence and capsule endoscopy: automatic detection of vascular lesions using a convolutional neural network

Introduction

Capsule endoscopy (CE) has revolutionized the approach to patients with suspected small intestine disease, allowing noninvasive inspection of this portion of the gastrointestinal tract. The clinical value of CE has been demonstrated in a wide array of diseases, including the evaluation of patients with suspected small bowel hemorrhage, diagnosis and monitoring of Crohn’s disease activity, and detection of protruding small intestinal lesions [1-4].

Obscure gastrointestinal bleeding (OGIB), either overt or occult, is responsible for 5% of all gastrointestinal hemorrhage cases. OGIB is currently the most frequent indication for CE. The source of bleeding is located in the small intestine in most cases [5,6]. A classification of CE findings according to their bleeding potential has been proposed by Saurin et al [7]. This classification defines CE findings as having no bleeding potential (P0), uncertain bleeding potential (P1), or high bleeding potential (P2). Findings with high bleeding potential include large ulcers, angiectasia, and varices. Vascular lesions are among the most commonly diagnosed lesions during the investigation of OGIB. Angiectasia is the most commonly found lesion with high bleeding potential among patients undergoing CE for OGIB [5]. These lesions are associated with chronic blood loss and are most commonly found in the elderly, patients with chronic kidney disease, cardiovascular disease and cirrhosis [8,9]. Other vascular lesions include enteric varices and red spots, which may be found in portal hypertension or diseases with systemic involvement [10-12]. In fact, OGIB is the most common indication for CE in patients with cirrhosis [13]. Furthermore, although the bleeding potential of mucosal red spots is uncertain, they are frequently found during CE for investigation of unexplained iron deficiency anemia and can occur as manifestations of several conditions, including portal hypertensive enteropathy and systemic vasculitis [11,12].

Evaluation of CE exams can be a burdensome task. Each CE video comprises approximately 50,000 frames, requiring an average of 30-120 min for reading [14]. Thus, this process is time-consuming for the clinical gastroenterologist. Moreover, mucosal lesions may be restricted to a small number of frames, increasing the risk of overlooking significant lesions.

The existence of large image databases and enhanced computational power have boosted the development of artificial intelligence (AI) tools for automatic image analysis. Among the different types of AI, convolutional neural networks (CNN) have delivered promising results in diverse fields of medicine [15-17]. Endoscopic imaging, and particularly CE, is one of the branches which can benefit the most from the development of CNN-based tools for the automatic detection of lesions [18]. These technological advances may increase diagnostic rates and optimize the reading process, including its time cost, which constitutes one of the main drawbacks of CE. Therefore, we aimed to create a CNN capable of automatically detecting and differentiating small intestinal vascular lesions of distinct bleeding potential, including red spots, angiectasias, and varices.

Patients and methods

Development of the CNN

From the collected pool of images (n=11588), 9525 contained normal enteric mucosa, 1026 had evidence of red spots (P1 lesions), and 1037 had angiectasia or varices (P2 lesions). This pool of images was split for the construction of training and validation image sets. The training dataset was composed by selecting 80% of the consecutively extracted images (n=9270). The remaining 20% were used as the validation dataset (n=2318). The validation dataset was used for assessing the performance of the CNN. A flowchart summarizing the study design and image selection for the development (training and validation) of the CNN is presented in Fig. 1.

Figure 1 Study flow chart for the training and validation phases CNN, convolutional neural network; CE, capsule endoscopy; AUROC, area under the receiver operating characteristic curve

To create the CNN, we used the Xception model with its weights trained on ImageNet (a large-scale image dataset aimed for use in development of object recognition software). To transfer this learning to our data, we kept the convolutional layers of the model. We removed the last fully connected layers and attached fully connected layers based on the number of classes we used to classify our endoscopic images. We used 2 blocks, each having a fully connected layer followed by a Dropout layer of 0.3 drop rate. Following these 2 blocks, we added a Dense layer with a size defined as the number of categories (n=3) to classify. The learning rate of 0.0001, batch size of 32, and the number of epochs of 100 were set by trial and error. We used Tensorflow 2.3 and Keras libraries to prepare the data and run the model. The analyses were performed using a computer equipped with a 2.1 GHz Intel^® Xeon^® Gold 6130 processor (Intel, Santa Clara, CA, USA) and a double NVIDIA Quadro^® RTX^™ 4000 graphic processing unit (NVIDIA Corporate, Santa Clara, CA, USA).

Statistical analysis

The primary outcome measures included sensitivity, specificity, precision, and the accuracy in differentiating between images containing normal mucosa, red spots and P2 lesions. In addition, we used receiver operating characteristic (ROC) curve analysis and area under the ROC curve (AUROC) to measure the performance of our model in the distinction between the 3 categories. The network’s classification was compared to the diagnosis provided by specialists’ analysis, the latter being considered the gold standard. In addition to its diagnostic performance, the computational speed of the network was determined using the validation image dataset by calculating the time required for the CNN to provide output for all images. For each image, the CNN calculated the probability for each of the 3 categories (normal mucosa, red spots and P2 lesions). A higher probability value translated into a greater confidence in the CNN prediction. The category with the highest probability score was outputted as the CNN’s predicted classification (Fig. 2). Sensitivities, specificities, and precisions are presented as mean ± standard deviation. ROC curves were graphically represented and AUROC calculated as mean and 95% confidence intervals (CI), assuming normal distribution of these variables. Statistical analysis was performed using Sci-Kit learn v0.22.2 [21].

Figure 2 Output obtained from the application of the convolutional neural network. The bars represent the probability estimated by the network. The finding with the highest probability was output as the predicted classification. A blue bar represents a correct prediction. Red bars represent an incorrect prediction N, normal mucosa; P1, red spots; P2, angiectasia and varices

Results

Construction of the network

A total of 1229 patients underwent CE and were enrolled in this study, from which 11,588 frames were extracted. The validation dataset comprised 2318 images (20% of the extracted frames). It was composed of 206 (8.9%) images with red spots, 207 (8.9%) images with P2 findings (angiectasia and varices), and 1905 (82.2%) images with normal mucosa. The CNN evaluated each image and predicted a classification (normal mucosa, red spots or P2 lesions) that was compared with the classification provided by the specialists. The network demonstrated its learning ability, with increasing accuracy as data were repeatedly input into the multi-layer CNN (Fig. 3).

Figure 3 Evolution of the accuracy of the convolutional neural network during training and validation phases, as the training and validation datasets were repeatedly input into the neural network

Overall performance of the network

The distribution of results is displayed in Table 1. Overall, the mean sensitivity and specificity of the CNN were 91.8±2.2% and 95.9±1.2%. The network provided accurate predictions in 94.4±3.7%. The positive predictive value was 91.3±3.7%. The negative predictive value was 95.7±2.1%.

Table 1 Confusion matrix of the automatic detection vs. expert classification

CNN performance for the detection and distinction of normal mucosa or enteric vascular lesions

We aimed to evaluate the CNN’s performance in the detection and distinction of enteric vascular lesions. The trained CNN had a sensitivity of 91.7%, specificity of 95.3%, and an accuracy of 94.1% for the detection of P1 lesions (red spots) (Table 2). The AUROC was 0.97. The network detected varices and angiectasia (P2 lesions) with a sensitivity, specificity and accuracy of 94.1%, 95.1% and 94.8%, respectively, and had an AUROC of 0.98. Classification as normal mucosa occurred with a sensitivity and specificity of 89.8% and 97.2% (Table 2), respectively, and an AUROC of 0.98. The ROC curves and respective AUROCs for detection of red spots, P2 lesions, and normal mucosa are represented in Fig. 4.

Table 2 CNN performance for detection and differentiation of red spots (P1) and P2 lesions

Figure 4 ROC analyses of the network’s performance in the detection of normal mucosa, P1 vascular lesions (red spots) and P2 vascular lesions (angiectasia and varices) ROC, receiver operating characteristic; AUC, area under the curve

Discussion

In this study, we developed an accurate deep learning tool for detection and differentiation of enteric lesions with distinct hemorrhagic potential. To the best of our knowledge, this is the first study to evaluate the performance of a CNN for detection of a wide range of vascular lesions with different bleeding potential. Our network reached high levels of performance in the detection of findings with uncertain and high bleeding potential. We believe that these results are promising for the development and introduction into clinical practice of tools for the automatic detection and classification of small intestine vascular lesions.

CE has revolutionized the etiologic investigation of the patient presenting with non-emergent OGIB, either overt or occult. The diagnostic yield of CE for OGIB is superior to the generality of other noninvasive diagnostic methods and is comparable to the much more invasive device-assisted double-balloon enteroscopy (DBE) [2]. The cost-effectiveness of CE in the setting of OGIB has been demonstrated [22,23]. Moreover, application of CE may synergistically enhance the diagnostic yield of deep enteroscopy techniques (e.g., DBE), thus selecting patients who may benefit from more invasive techniques with therapeutic potential [2,24].

The accurate and timely detection of small intestine lesions in CE is essential. Saurin et al have created a useful and pragmatic classification of the bleeding potential for lesions detected on CE [7]. Vascular lesions, including angiectasias and varices, present high bleeding risk (P2). Red spots have uncertain/intermediate clinical significance [7]. Nevertheless, they are frequently found in patients with occult OGIB without other high-risk lesions and may be a sign of systemic diseases [12]. However, the pleomorphism of these lesions as well as their frequent small size may increase the risk of missing lesions. To the best of our knowledge, this is the first CNN to specifically identify and distinguish red spots from other vascular lesions and normal mucosa.

Reading CE exams is a time-consuming task, and significant lesions may be restricted to a small number of frames, thus increasing the risk of overlooking significant lesions [25,26]. The application of AI tools to CE may allow these drawbacks to be overcome. Assistance of deep learning methods such as CNNs may improve detection of these lesions while shortening the time required for reading the images [27,28]. Recent application of these technologies has confirmed the high diagnostic performance of CNN-based models for small-bowel CE, including for the detection of ulcers and erosions, protruding lesions, celiac disease, luminal blood content, and angiectasia [25,29-33].

This is the first study reporting the performance of a CNN-based model for the detection of lesions with distinct hemorrhagic potential on CE images. To date, the existing studies regarding automatic detection of vascular lesions have focused on the detection of angiectasia. Noya et al reported the development of a CNN for detection of angiectasia with a sensitivity of 90%, a specificity of 96.8% and an AUROC of 0.93 [34]. Leenhardt et al developed a CNN capable of detecting angiectasia with a sensitivity of 100% and specificity of 96% [31]. However, that study was performed with frames extracted from the French national CE image database, which included only clean images. Therefore, their results may not be generalizable because of variations in bowel preparation and image artifacts. In 2019, Tsuboi et al developed a CNN-based system for automatic detection of angiectasia [35]. Their system proved to have a high diagnostic yield for the detection of angiectasia, with a sensitivity of 98.8% and specificity of 98.4%. However, this study failed to address the detection of lesions according to their hemorrhagic potential, which is clinically important, as lesions with different bleeding potential have distinct rebleeding rates [36]. More recently, a CNN model developed by Otani et al reported an AUROC of 0.95 for the detection of vascular lesions (angiectasias and venous malformations) [37].

A significant number of patients presenting with OGIB are concomitantly treated with antiplatelet or anticoagulant drugs for comorbid cardiovascular disease [38]. Special attention is required for these patients, as these therapies are associated with increased risk of mucosal injury or bleeding from preexistent vascular lesions [38,39]. The detection of these lesions in CE is important for the management of OGIB in this subset of patients. In this study, we developed a CNN-based model with high diagnostic yield for the detection of red spots (P1) and P2 lesions (angiectasias and varices). Indeed, this is the first published CNN to differentiate vascular lesions, rather than simply identifying them, as reported in previously published CNNs regarding vascular lesions. Furthermore, our CNN is able to stratify hemorrhagic risk, by accurately classifying frames according to Saurin’s classification. The development of sensitive AI tools for the automatic detection of vascular lesions in CE images may improve the diagnostic yield of CE for these lesions, thus decreasing the number of negative CE exams in the setting of OGIB. Moreover, the development of algorithms capable of foretelling the hemorrhagic potential of vascular lesions may help stratify patients who require further evaluation. This might translate into future gains regarding adequate management of healthcare resources.

This work has several highlights. First, to our knowledge, this is the first study to evaluate the performance of CNN for the detection of several vascular lesions. Moreover, we tested the ability of our model in the detection and discrimination of lesions with different bleeding potential. Second, our algorithm demonstrated high levels of performance in the detection and differentiation of such lesions. The sensitivity, specificity and AUROC for the detection of red spots and P2 lesions (angiectasias/varices) were, respectively, 91.7%, 65.3% and 0.97, and 94.1%, 95.1% and 0.98. Third, the architecture of our network demonstrated a high image processing performance, with approximate reading rates of 145 frames/sec. This performance is superior to most studies regarding automatic detection of lesions, including vascular abnormalities [25,29-31,35]. We believe that this performance may, in the near future, translate into shorter reading times, thus overcoming one of the main drawbacks of CE. Further prospective multicentric studies are required to assess whether AI-assisted CE image reading translates into enhanced time efficiency compared to conventional reading.

This study had several limitations. First, our study focused on patients evaluated in a single center and was conducted in a retrospective manner. Thus, these promising results must be confirmed by robust prospective multicenter studies before application to clinical practice. Second, this model was developed using PillCam SB3. Therefore, our results may not be generalizable for other CE systems. Third, our system was developed using still frames. Assessing the performance of this technology using full-length videos is required before the clinical implementation of this model. Third, although our network demonstrated high processing speed, we did not assess whether CNN-assisted image review reduces the reading time compared to conventional reading. Finally, the number of included patients was relatively small. Moreover, the number of images in the validation dataset was small, thus limiting the interpretation of our results. Therefore, prospective investigation in larger patient sets is required to confirm these results before the introduction of these tools into clinical practice.

The implementation of AI tools in routine clinical practice is expected to grow in the near future. CE is a fertile ground for the development of deep learning-based tools for enhanced image processing. These tools may help reduce CE reading times, thus overcoming one of its main downsides, as well as improving the diagnostic accuracy of CE for multiple lesions.

In conclusion, we developed a CNN-based model capable of detecting and discriminating vascular lesions with distinct hemorrhagic potential. Our model achieved high levels of accuracy and excellent image processing performance. We believe that our results may help lay the foundations for the widespread application of AI technology in the field of CE.