1. Introduction
Inflammatory bowel disease is a chronic inflammatory condition characterized by relapsing-remitting inflammation of the gastrointestinal tract. It includes two main entities: ulcerative colitis, which affects the colon, and Crohn’s disease, which can involve any part of the gastrointestinal tract [
1]. The estimated prevalence of inflammatory bowel disease is up to 0.8% in countries such as the UK [
2]. The pathogenesis of this condition is still not well understood in both entities, and there is overlap between them [
3]. In pathogenesis, both host immune response and microbial factors are involved, including alterations of the epithelial barrier [
4], dysregulation of immune cells, dysregulation of secreted mediators, microbes, and genetic susceptibility [
5] (
Table 1).
Ulcerative colitis is an inflammatory disease limited to the colon that usually affects the rectum and extends to the proximal side. The prevalence of ulcerative colitis is estimated to be 5 million cases worldwide [
45]. The onset of the disease is usually gradual and progressive over time, and is typically accompanied by diarrhea. Systemic symptoms include weight loss, fatigue, and fever [
5,
45]. The disease severity ranges from mild, to moderate, and severe, and the Mayo scoring system/Disease Activity Index (DAI) can be used to assess disease severity and monitor response to therapy [
46,
47]. The variables of the Mayo score are stool pattern, most severe rectal bleeding of the day, endoscopic findings, and global assessment by clinician [
48].
Acute complications of ulcerative colitis include severe bleeding that occurs in up to 10% of patients, fulminant colitis and toxic megacolon, and perforation [
5,
45]. Ulcerative colitis is primary a disease of the colon. However, extraintestinal manifestations are also present: musculoskeletal (arthritis and arthropathy), eye (uveitis and episcleritis), skin (erythema nodosum and pyoderma gangrenosum), hepatobiliary (primary sclerosing cholangitis, fatty liver and autoimmune liver disease), hematopoietic/coagulation (thromboembolism), and pulmonary [
5,
45,
49,
50,
51,
52,
53,
54].
The diagnosis of ulcerative colitis is based on the presence of chronic diarrhea for > 4 weeks and evidence of colon inflammation on histological biopsy [
45]. Since these characteristics are not specific, other diseases must be excluded before reaching the diagnosis of ulcerative colitis, including Crohn’s disease, infection colitis, radiation colitis, diverticulitis, diversion colitis, solitary rectal ulcer syndrome, graft-versus-host disease, and medication-associated colitis [
5]. Of note, patients with ulcerative colitis have a higher risk of developing dysplasia and colorectal cancer [
5]. The extent of the affected area (patients with pancolitis) and duration of the disease are the two major risk factors associated with the development of neoplasia [
55].
The aim of treatment in patients with active ulcerative colitis is to achieve clinical and endoscopic remission. Mesalamine (mesalazine) is a 5-aminosalicylic acid derivative is initially used. If there is no improvement, glucocorticoids are used (for example, budesonide, or prednisone). Failure to respond and refractory disease may require systemic glucocorticoids and biological agents, such as anti-tumor necrosis (TNF) agents, anti-α4β7-integrin (vedolizumab), anti-interleukin antibody-based therapy, sphingosine-1-phosphate (S1P) receptor modulators, or small molecules (tofacitinib, an small-molecule Janus kinase inhibitor) [
56,
57,
58,
59,
60,
61,
62,
63,
64,
65,
66].
Colorectal cancer (CRC) is a common and lethal disease, with an annual incidence of approximately 153,000 cases [
67]. The risk of CRC depends on environmental and genetic factors such as inflammatory bowel disease [
68]. The diagnosis is usually made by colonoscopy. The management of localized disease is surgical resection and adjuvant chemotherapy [
69].
Convolutional neural networks for deep learning image classification is an application of digital pathology and artificial intelligence in translational medicine and clinical practice [
70]. The deep learning workflow includes data preprocessing, network building, training, network performance improvement by tuning hyperparameters or running multiple trials, and visualization and verification of network behavior during and after training [
71].
This study used several convolutional neural networks to classify images of ulcerative colitis and differentiate between colonic control and colorectal cancer (adenocarcinoma). In addition, the protein expression of a new immuno-oncology maker was explored in the ulcerative colitis cases.
2. Materials and Methods
Hematoxylin and eosin (H&E) histological slides of 35 ulcerative colitis patients were retrieved from our previous publication [
72]. All patients were started on 5-ASA (mesalazine) treatment with or without probiotics. A more aggressive treatment (prednisolone as the first choice) was used if the initial treatment failed to induct remission state (UC-DAI score, 1–2) or if the disease relapsed as defined by symptoms (UC-DAI > 2, with bloody stool) as well as by laboratory data and/or colonoscopy [
72]. After a follow-up of 2 years, 35 cases were classified into two groups: 22 cases of mesalazine-responsive ulcerative colitis (clinically denominated as “benign”) and 13 cases of steroids-requiring ulcerative colitis (clinically characterized by a more “aggressive” behavior).
Other recorded clinicopathological characteristics included age at biopsy, sex, biopsy location, Baron score for endoscopic grading of ulcerative colitis [
73,
74,
75], and histopathological Geboes score for ulcerative colitis [
76] (
Table 2). This project was in accordance with the World Medical Association (WMA) Declaration of Helsinki on ethical principles for medical research involving human subjects (IRB 13R-119).
Immunohistochemistry targeting LAIR1 was performed using a Bond-Max fully automated immunohistochemistry and in-situ hybridization staining system following the manufacturer’s instructions (Leica Biosystems K.K., Tokyo, Japan). For immunohistochemistry using DAB chromogen and hematoxylin counterstain, polymer detection system DS9800 was used (Leica Byosystems). The primary antibody targeted leukocyte-associated immunoglobulin-like receptor (LAIR1/CD305) created by Dr. Giovanna Roncador of Monoclonal Antibodies Core Unit, located a Spanish National Cancer Research Center (also known as CNIO: Centro Nacional de Investigaciones Oncológicas). The LAIR1 is a rat monoclonal antibody, clone JAVI82A, antigen used: RBL-1-LAIR1-MYC-DDK transfected cells and last booster with LAIR1 recombinant (Gln22-His163, with a C-terminal 6-His tag); isotype IgG2a; reactivity, human; localization, membrane.
TOX2 antibody targeted TOX High Mobility Group Box Family Member 2, and was also developed by CNIO. Properties: clone name TOM924D, rat monoclonal, IgG2b K, antigen HIS-SUMO-hTOX2-Strep-tag2 full length protein, human reactivity, nucleus localization.
The slides were scanned using a NanoZoomer S360 virtual slide scanner (#C13220-01, Hamamatsu Photonics, Hamamatsu, Japan). Digital images were visualized using NDP.view2 software (#U12388-01, Hamamatsu Photonics), and each intestinal biopsy was exported into a jpeg file at 200× magnification and 150 dpi.
Images were cropped at 243×243 size, reviewed by the pathologist (J.C.), and unproductive images were excluded from the analysis. The filtering criteria were as follows: (1) cropped images of not 243×243 size; (2) images less than 5-31 KB that usually do not contain tissue; (3) images that did not contain diagnostic areas based on histopathological criteria; (4) images with artifacts such as broken and folded tissue and wrongly stained. The series also included 18 cases of colorectal cancer (adenocarcinoma) and 21 patients of colon control who underwent diagnostic biopsies and surgical resection specimens.
A convolutional neural network (CNN) was designed based on transfer learning from ResNet-18; and trained to classify the 3 types of images, ulcerative colitis (n = 9,281), colon control (n = 12,246), and colorectal cancer (n = 63,725). The data were partitioned into a training set (70% of the images) to train the network, a validation set (10%) to test the performance of the network during training, and a test set (20%) as a holdout (new data) to test the performance on new data. The order of the images was randomized to ensure that the learning of classes evenly.
Data normalization was applied to the input images as previously described [
71]. The code was run in Matlab programing language and numeric computing environment (R2023b, update 9 released 30 Jul 2024, MathWorks, Tokyo, Japan), as we have recently described [
71]. The input size was 224-by-224 (224×224×3). In summary, the code loaded the pretrained CNN, replaced the final layers, trained the network, made predictions, assessed the accuracy, and deployed the results. The design and training parameters of the CNN are listed in
Table 3.
Advanced explainable artificial intelligence for computer vision was performed to understand the classification decisions by the deep learning network using the gradient-weighted class activation mapping (Grad-CAM) heatmap technique.
The performance of the ResNet-18-based network was compared to other pre-trained networks, including AlexNet, DenseNet-201, EfficientNet-b0, GoogLeNet, Inception-v3, MobileNet-v2, NASNet-Large, NASNet-Mobile, ResNet-18, ResNet-50, ResNet-101, ShuffleNet, VGG-16, VGG-19, and Xception (
Table 4).
Figure 1.
The 3 types of samples included in this study: ulcerative colitis (left), colorectal cancer (adenocarcinoma, middle), and colon control (right). Original magnification 100×.
Figure 1.
The 3 types of samples included in this study: ulcerative colitis (left), colorectal cancer (adenocarcinoma, middle), and colon control (right). Original magnification 100×.
Figure 2.
Cropped images of ulcerative colitis. Original magnification 200×. Images were cropped at 224 × 224 size. Ulcerative colitis is an idiopathic chronic inflammation that affects the colon mucosa. This disorder characteristically affects the rectum and extends toward proximal sections of the colon in a continuous manner. Microscopically, there are signs of active chronic colitis when untreated. Chronicity includes distorted architecture of the crypts, such as atrophy, irregular spacing, shortening, and branching; inflammation of the lamina propria with basal lymphoplasmacytosis, and Panet cell metaplasia or hyperplasia. Disease activity is confirmed by neutrophil infiltration of the muscosa, cryptitis, crypt abscess, or ulceration. Typically, inflammation is limited to the mucosa and submucosa, and granulomas and fissuring ulcers are absent. Well-established disease can be associated with dysplasia of the epithelium, either low-grade, or high-grade. The most commonly used score to evaluate the histological features is the Geboes Score [
89,
90,
91,
92,
93].
Figure 2.
Cropped images of ulcerative colitis. Original magnification 200×. Images were cropped at 224 × 224 size. Ulcerative colitis is an idiopathic chronic inflammation that affects the colon mucosa. This disorder characteristically affects the rectum and extends toward proximal sections of the colon in a continuous manner. Microscopically, there are signs of active chronic colitis when untreated. Chronicity includes distorted architecture of the crypts, such as atrophy, irregular spacing, shortening, and branching; inflammation of the lamina propria with basal lymphoplasmacytosis, and Panet cell metaplasia or hyperplasia. Disease activity is confirmed by neutrophil infiltration of the muscosa, cryptitis, crypt abscess, or ulceration. Typically, inflammation is limited to the mucosa and submucosa, and granulomas and fissuring ulcers are absent. Well-established disease can be associated with dysplasia of the epithelium, either low-grade, or high-grade. The most commonly used score to evaluate the histological features is the Geboes Score [
89,
90,
91,
92,
93].
Figure 3.
Cropped images of colorectal cancer (adenocarcinoma). Original magnification 200×. Images were cropped at 224 × 224 size. Adenocarcinoma of the colon is a glandular neoplasm that accounts for approximately 98% of all colonic cancers. Patients with inflammatory bowel disease, polyposis, and Lynch syndrome [
94,
95] are at higher risk of developing colorectal cancer. Most cases display well- or moderate differentiation of carcinoma glands accompanied by marked growth of fibrous connective tissue known as desmoplasia [
96,
97]. The glands can show a cribriform pattern and are filled with necrotic debris. Adenocarcinomas are characterized by epithelial cells with stretched and stratified nuclei, which create complex glandular structures. The nuclei exhibit polymorphism and loss of polarity. The tumor immune microenvironment exhibits variable infiltration of inflammatory cells. There are several recognized subtypes, including adenoma-like, adenosquamous, mucinous, micropapillary, signet-ring, serrated, and sarcomatoid [
89,
97,
98,
99,
100,
101,
102,
103].
Figure 3.
Cropped images of colorectal cancer (adenocarcinoma). Original magnification 200×. Images were cropped at 224 × 224 size. Adenocarcinoma of the colon is a glandular neoplasm that accounts for approximately 98% of all colonic cancers. Patients with inflammatory bowel disease, polyposis, and Lynch syndrome [
94,
95] are at higher risk of developing colorectal cancer. Most cases display well- or moderate differentiation of carcinoma glands accompanied by marked growth of fibrous connective tissue known as desmoplasia [
96,
97]. The glands can show a cribriform pattern and are filled with necrotic debris. Adenocarcinomas are characterized by epithelial cells with stretched and stratified nuclei, which create complex glandular structures. The nuclei exhibit polymorphism and loss of polarity. The tumor immune microenvironment exhibits variable infiltration of inflammatory cells. There are several recognized subtypes, including adenoma-like, adenosquamous, mucinous, micropapillary, signet-ring, serrated, and sarcomatoid [
89,
97,
98,
99,
100,
101,
102,
103].
Figure 4.
Cropped images of colon control. Original magnification 200×. Images were cropped at 224 × 224 size. The colonic mucosa has the function of absorbing water and electrolytes, which is performed by absorptive columnar cells, and producing of mucus to lubricate, which is produced by globet cells. The mucosa is composed of the epithelium, lamina propria, and muscularis mucosa. The epithelium invaginates and forms glands (crypts), where at its base there is also the presence of enteroendocrine, Paneth cells, and stem cells. The lamina propria is rich in capillaries and lymphatics. Loose connective tissue and nerve plexuses are found in the submucosa. The muscularis propria has an inner circular layer and an outer longitudinal layer, and within them, the Auerbach nerve plexus is located. The outer layers are the subserosa and serosa [
89,
104,
105].
Figure 4.
Cropped images of colon control. Original magnification 200×. Images were cropped at 224 × 224 size. The colonic mucosa has the function of absorbing water and electrolytes, which is performed by absorptive columnar cells, and producing of mucus to lubricate, which is produced by globet cells. The mucosa is composed of the epithelium, lamina propria, and muscularis mucosa. The epithelium invaginates and forms glands (crypts), where at its base there is also the presence of enteroendocrine, Paneth cells, and stem cells. The lamina propria is rich in capillaries and lymphatics. Loose connective tissue and nerve plexuses are found in the submucosa. The muscularis propria has an inner circular layer and an outer longitudinal layer, and within them, the Auerbach nerve plexus is located. The outer layers are the subserosa and serosa [
89,
104,
105].
4. Discussion
Ulcerative colitis is a chronic inflammatory bowel disease that primary affects the mucosal layer of the colon recurrently. It usually affects the rectum and extends continuously toward the proximal segments of the colon [
5,
106]. Patients usually present with diarrhea that may include blood, with a gradual and progressive onset of symptoms [
5,
45,
106]. Disease evaluation includes history, laboratory studies, endoscopy, and biopsy. Several factors affect the disease course, including age at diagnosis, mucosal healing, extension of colitis, and smoking. Chronic complications include stricture, dysplasia, and colorectal cancer [
49].
The endoscopic findings are nonspecific and include loss of vascular marking, granular mucosa, petechiae, exudates, edema, erosions, friability, ulcerations, and bleeding [
107]. The endoscopic biopsies show neutrophilic infiltration with cryptitis, crypt abscesses, and ulcerations when the disease is active. Chronic disease involves crypt architectural alterations, chronic inflammation of the lamina propria with lymphoplasmocytosis, and Paneth cell metaplasia or hyperplasia [
89,
90,
104]. This study analyzed images of ulcerative colitis using convolutional neural network (CNN).
A CNN is a deep learning machine learning technique trained using a large number of images. CNN was used to analyze whole slide images (WSI) from hematoxylin & eosin (H&E) microscope pathology slides of ulcerative colitis, colon control, and colorectal cancer (adenocarcinoma). Artificial intelligence (AI) is the process of being capable of simulating human intelligence, to mimic the cognitive functions of learning and problem solving. AI can be divided into artificial general intelligence, or strong AI, that has generalized human cognitive abilities [
108,
109,
110]; and narrow AI, or weak AI, that focuses on a specific task of human intelligence [
111,
112]. There are two types of narrow AI: rule-based AI is provided with machine rules to follow, and example-based AI is provided with examples to learn from [
113]. This study was a histological analysis using narrow AI to classify the gut images. The slides were digitalized using a slide scanner that converted glass slides into digital data. In this study, both endoscopic biopsy and surgical rejection specimens were used. The CNN could differentiate the 3 types of diagnoses with high performance (99%). However, the differential diagnosis of ulcerative colitis must be considered.
The evaluation and establishment of a diagnosis of ulcerative colitis need to exclude other causes of colitis using history, laboratory studies, endoscopic images, and colon biopsies. In the history, other causes of colitis should be excluded, including parasitic infections, sexually transmitted infections (Neisseria gonorrhoeae and herpes simplex virus), atherosclerotic disease (chronic colonic ischemia), abdominal/pelvic radiation, and use of non-steroidal anti-inflammatory drugs. The stool should also be tested for C. trachomatis, N. gonorrhoeae, HSV, and Treponema pallidum. The endoscopic and histological findings of ulcerative colitis are not specific. However, differentiating ulcerative colitis from Crohn’s disease is important because of its different prognoses and treatments [
114]. A comprehensive review and update on ulcerative colitis, including an extended differential diagnosis, was described by Gajendran M. et al. [
115]. Among the relevant diseases, neoplasm, Crohn’s disease, and Celiac disease were highlighted [
115]. In this study, the image classification comprised cases of ulcerative colitis and colorectal cancer (adenocarcinoma). We recently performed histological image classification of Celiac disease, small intestine control, unspecific duodenal inflammation, and Crohn’s disease [
71]. Narrow artificial intelligence (AI) is designed to perform tasks that typically require human intelligence; however, it operates within limited constraints and is task-specific [
71]. In future, integrated analysis can performed.
Computer vision provides a series of algorithms for visual inspection, object detection, and tracking, and feature detection, extraction, and matching. Pretrained object detection can be performed using the YOLO, SSD, and ACF algorithms. Semantic and instance segmentation using U-Net, SOLO, and the Mask R-CNN [
116]. Image classification can be performed using vision transformers, such as ViT. In this study, we employed pretrained image classification neural networks. These pretrained networks have already learned to extract the characteristics of natural images. Therefore, we used them as a starting point to learn the new task of classifying ulcerative colitis, colonic control, and colorectal cancer (adenocarcinoma). Several CNNs were used, and the most accurate ones were DenseNet-201 (99.30% accuracy), ResNet-50 (99.14%), Inception-v3 (99.13%), ResNet-101 (99.1%), and ResNet-18 (99%). The most relevant features of a CNN are accuracy, speed, and size. In this study, ResNet-18 was the most convenient CNN to use due to its accuracy and relative prediction time using GPU. NASNet-Large provided good performance; however, despite having high accuracy, the prediction time was the slowest.
Explainable AI (XAI) allows humans to understand and trust results by providing clear and understandable explanations [
117,
118,
119,
120]. In computer vision, a common method to demonstrate XAI techniques is to overlay an explanation on the image using an explanation heatmap. In this study, grad-CAM was used [
121,
122]. The other XAI methods are LIME and SHAP [
123]. Grad-CAM uses a gradient-weighted class activation mapping technique to understand why a deep learning network makes classification decisions. As shown in the results section, the Grad-CAM analysis confirmed that the CNN focused on the correct area of the images to perform classification. Interestingly, in the discordant cases, the Grad-CAM analysis revealed incorrect focus by the CNN. In addition, some discordant cases were due to nondiagnostic images.
In the last 5 years there have been several reports on ulcerative colitis and deep learning. Most applications have focused on endoscopic image analysis [
124,
125,
126,
127,
128,
129,
130,
131]. However, some research has been published regarding histological images. Dawid Rymarczk et al. used deep learning models to analyze the histological disease activity in Crohn’s disease and ulcerative colitis [
132]. Niels Vande Casteele et al. used a deep learning algorithm to identify eosinophils in colonic biopsies of active ulcerative colitis [
133]. Jun Ohara et al. applied deep learning to detect goblet cell mucus [
134]. Laurent Peyrin-Biroulet used an AI algorithm to measure the Nancy index in histological images of patients with ulcerative colitis [
135]. Our study is different from previous publications because made diagnosis of ulcerative colitis against colon control and identified images of colorectal cancer. To the best of our knowledge, this type of study has not been conducted before now.
LAIR1 protein functions an inhibitory receptor that is expressed in monocytes, natural killer cells, T and B-lymphocytes. LAIR1 is also expressed in macrophages and regulates their activation [
136]. In lymphoma, LAIR1 expression by tumor-associated macrophages has been described [
137]. In ulcerative colitis, the expression of LAIR1 has been recently described by Mina Hassan-Zahraee et al. [
138]. However, to the best of our knowedge, LAIR1 has not been described in other inflammatory bowel diseases. In our study, LAIR1 was expressed by cells of the immune microenvironment, and the CNN managed to classify between steroid-requiring (SR) and nonsteroid requiring (non-SR) ulcerative colitis. Therefore, LAIR1 is a promising immuno-oncology marker in ulcerative colitis.
TOX2 (TOX High Mobility Group Box Family Member 2) is a transcription factor related to T cell exhaustion [
139], which is a broad term used to describe the T cell functions in situation of chronic antigen stimulation, such as inflammatory bowel disease and response to tumors [
140]. Evidence indicates that TOX2 is expressed in T follicular helper (TFH) cells, similar to PD-1 marker, and that would suppress CD4 cytotoxic T cell differentiation [
141]. TOX2 has been related to the survival of patients with acute myeloid leukemia [
140], and to the pathogenesis of atopic dermatitis [142]. Because of the relationship with PD-1, TOX2 is a promising immuno-oncology marker in ulcerative colitis as well.
Figure 5.
Network performance during training and validation. The data were partitioned into a training set (70% of the images) to train the network and a validation set (10%) to test the performance of the network during the training; and a test set (20%) was used as a holdout to test the performance of the trained network on new data. This figure shows the accuracy and loss during the training (70%) and validation (10%) sets. The CNN was based on transfer learning from ResNet-18).
Figure 5.
Network performance during training and validation. The data were partitioned into a training set (70% of the images) to train the network and a validation set (10%) to test the performance of the network during the training; and a test set (20%) was used as a holdout to test the performance of the trained network on new data. This figure shows the accuracy and loss during the training (70%) and validation (10%) sets. The CNN was based on transfer learning from ResNet-18).
Figure 6.
Confusion matrix of the test dataset (new data). The data were partitioned into a training set (70% of the images) to train the network and a validation set (10%) to test the performance of the network during the training; and a test set (20%) was used as a holdout to test the performance of the trained network on new data. ADK, adenocarcinoma (colorectal cancer); CC, colon control; UC, ulcerative colitis.
Figure 6.
Confusion matrix of the test dataset (new data). The data were partitioned into a training set (70% of the images) to train the network and a validation set (10%) to test the performance of the network during the training; and a test set (20%) was used as a holdout to test the performance of the trained network on new data. ADK, adenocarcinoma (colorectal cancer); CC, colon control; UC, ulcerative colitis.
Figure 7.
Explanation of network predictions using Grad-CAM. Grad-CAM was used to visualize which parts of the image were important for the classification decision (diagnosis) of the network. The most relevant parts are highlighted in red (jet colormap). The prediction scores for each diagnosis are shown below each hematoxylin and eosin (H&E) image.
Figure 7.
Explanation of network predictions using Grad-CAM. Grad-CAM was used to visualize which parts of the image were important for the classification decision (diagnosis) of the network. The most relevant parts are highlighted in red (jet colormap). The prediction scores for each diagnosis are shown below each hematoxylin and eosin (H&E) image.
Figure 8.
Grad-CAM analysis of incorrectly classified images. Grad-CAM analysis was used to visualize which parts of the image were important for the classification decision (diagnosis) of the network. In most cases, the classification error was due to a wrong focalized area in the image by the neural network or the image was not diagnostic from a histopathological point of view.
Figure 8.
Grad-CAM analysis of incorrectly classified images. Grad-CAM analysis was used to visualize which parts of the image were important for the classification decision (diagnosis) of the network. In most cases, the classification error was due to a wrong focalized area in the image by the neural network or the image was not diagnostic from a histopathological point of view.
Figure 9.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 79.53.
Figure 9.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 79.53.
Figure 10.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive using LAIR1 immunohistochemistry. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 88.31%.
Figure 10.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive using LAIR1 immunohistochemistry. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 88.31%.
Figure 11.
LAIR1 immunohistochemistry in the ulcerative colitis dataset, and classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative colitis cases. Overall, the inflammatory component was higher in steroid-requiring cases.
Figure 11.
LAIR1 immunohistochemistry in the ulcerative colitis dataset, and classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative colitis cases. Overall, the inflammatory component was higher in steroid-requiring cases.
Figure 12.
Examples of cropped images of LAIR1 immunohistochemistry in the ulcerative colitis dataset. The cropped images were used as input data in the CNN, which managed to classify between steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative cases. Overall, the inflammatory component was higher in steroid-requiring cases.
Figure 12.
Examples of cropped images of LAIR1 immunohistochemistry in the ulcerative colitis dataset. The cropped images were used as input data in the CNN, which managed to classify between steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative cases. Overall, the inflammatory component was higher in steroid-requiring cases.
Figure 13.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive using TOX2 immunohistochemistry. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 85.62%.
Figure 13.
Confusion matrix of the test dataset for the classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive using TOX2 immunohistochemistry. The test dataset included new data (holdout, 20%). The analysis was based on transfer learning from ResNet-18. The accuracy was 85.62%.
Figure 14.
TOX2 immunohistochemistry in the ulcerative colitis dataset, and classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative colitis cases. Overall, the TOX2-positive inflammatory component was higher in mesalazine-responsive cases.
Figure 14.
TOX2 immunohistochemistry in the ulcerative colitis dataset, and classification of steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative colitis cases. Overall, the TOX2-positive inflammatory component was higher in mesalazine-responsive cases.
Figure 15.
Examples of cropped images of TOX2 immunohistochemistry in the ulcerative colitis dataset. The cropped images were used as input data in the CNN, which managed to classify between steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative cases. Overall, the TOX2-positive inflammatory component was higher in mesalazine-responsive cases.
Figure 15.
Examples of cropped images of TOX2 immunohistochemistry in the ulcerative colitis dataset. The cropped images were used as input data in the CNN, which managed to classify between steroid-requiring and nonsteroid-requiring/mesalazine-responsive ulcerative cases. Overall, the TOX2-positive inflammatory component was higher in mesalazine-responsive cases.
Table 1.
Pathogenesis of inflammatory bowel disease.
Table 1.
Pathogenesis of inflammatory bowel disease.
Mechanisms |
Key players |
Dysregulation of the epithelial barrier |
Alterations of the mucus, increased number of bacteria within the mucus, and increased intestinal permeability [6,7,8] |
Dysregulation of immune cells |
Increased recruitment and activation of immune cell, including myeloid inflammatory cells, natural killer cells, T cells, B cells, plasma cells, neutrophils, and other leukocytes [9,10,11,12,13,14,15,16,17]. |
Dysregulation of immune regulators and inflammatory cytokines |
CD4+T lymphocytes, interferon (IFN)-gamma, Th1, Th2, Th17, FOXP3+regulatory T lymphocytes (Tregs), IL-10, TGFB, CD8+cytotoxic T lymphocytes [18,19,20,21,22,23,24,25] |
Microbes |
Alterations in the diversity and density of bacteria [26,27,28,29], specific microbial components, intestinal viruses [9,30,31,32], and fungi [33,34,35] |
Genetic susceptibility |
Over 240 different susceptibility loci, NOD2, ATG16L1, NADPH, and immune-related (Th17/IL-23, IL-10, TNFSF15, cytokine, adaptive immune response and epithelial pathways) [14,36,37,38,39,40,41,42,43,44] |
Table 2.
Clinicopathological characteristics of ulcerative colitis.
Table 2.
Clinicopathological characteristics of ulcerative colitis.
Variable No. (%) |
Mesalazine -responsive |
Steroid -requiring |
All cases |
P value |
Number of patients |
22 |
13 |
35 |
|
Age (mean ±STD) |
43.7 ±13.6 |
29.5 ±17.6 |
38.4 ±16.4 |
0.012 |
Sex Male |
14/22 (63.6) |
6/13 (46.2) |
20/35 (57.1) |
0.481 |
Colon biopsy location |
|
|
|
|
Ascending |
0/22 (0) |
1/13 (7.7) |
1/35 (2.9) |
0.009 |
Transverse |
0/22 (0) |
2/13 (15.4) |
2/35 (5.7) |
|
Descending |
2/22 (9.1) |
3/13 (23.1) |
5/35 (14.3) |
|
Sigmoid |
2/22 (9.1) |
3/13 (23.1) |
5/35 (14.3) |
|
Rectum |
18/22 (81.8) |
4/13 (30.8) |
22/35 (62.9) |
|
Endoscopic Baron score |
|
|
|
|
1 |
13/22 (59.1) |
2/13 (15.4 |
15/35 (42.9) |
0.009 |
2 |
9/22 (40.9) |
8/13 (61.5) |
17/35 (48.6) |
|
3 |
0/22 (0) |
3/13 (23.1) |
3/35 (8.6) |
|
Histologic Geboes score |
|
|
|
|
1 |
2/22 (9.1) |
0/13 (0) |
2/35 (5.) |
0.101 (0.007*) |
2 |
13/22 (59.1) |
4/13 (30.8) |
17/35 (48.6) |
|
3 |
5/22 (22.7) |
3/13 (23.1) |
8/35 (22.9) |
|
4 |
2/22 (9.1) |
5/13 (38.5) |
7/35 (20) |
|
5 |
0/22 (0) |
1/13 (7.7) |
1/35 (2.9) |
|
Table 3.
Design and training parameters of the convolutional neural network.
Table 3.
Design and training parameters of the convolutional neural network.
ResNet-18-based CNN |
Training (70%) |
Validation (10%) |
Training options |
Input type: image Output type: classification Number of layers: 71 Number of connections: 78 |
Observations: 59677 Classes: 3 Ulcerative colitis: 6497 Colorectal cancer: 44608 Colon control: 8572 |
Observations: 8525 Classes: 3 Ulcerative colitis: 928 Colorectal cancer: 6372 Colon control: 1225 |
Solver: sgdm Initial learning rate: 0.001 MiniBatch size: 128 MaxEpochs: 5 Validation frequency: 50 Iterations: 2330 Iterations per epoch: 466 |
Table 4.
Pretrained Neural Networks.
Table 4.
Pretrained Neural Networks.
Name [Reference] |
Model Name Argument |
Depth |
Size (MB) |
Parameters (Millions) |
Image Input Size |
AlexNet [77] |
“alexnet” |
8 |
227 |
61 |
227-by-227 |
DenseNet-201 [78] |
“densenet201” |
201 |
77 |
20 |
224-by-224 |
EfficientNet-b0 [79] |
“efficientnetb0” |
82 |
20 |
5.3 |
224-by-224 |
GoogLeNet [80,81] |
“googlenet” |
22 |
27 |
7 |
224-by-224 |
|
“googlenet-places365” |
|
|
|
|
Inception-v3 [82] |
“inceptionv3” |
48 |
89 |
23.9 |
299-by-299 |
MobileNet-v2 [83] |
“mobilenetv2” |
53 |
13 |
3.5 |
224-by-224 |
NASNet-Large [84] |
“nasnetlarge” |
* |
332 |
88.9 |
331-by-331 |
NASNet-Mobile [84] |
“nasnetmobile” |
* |
20 |
5.3 |
224-by-224 |
ResNet-18 [85] |
“resnet18” |
18 |
44 |
11.7 |
224-by-224 |
ResNet-50 [85] |
“resnet50” |
50 |
96 |
25.6 |
224-by-224 |
ResNet-101 [85] |
“resnet101” |
101 |
167 |
44.6 |
224-by-224 |
ShuffleNet [86] |
“shufflenet” |
50 |
5.4 |
1.4 |
224-by-224 |
VGG-16 [87] |
“vgg16” |
16 |
515 |
138 |
224-by-224 |
VGG-19 [87] |
“vgg19” |
19 |
535 |
144 |
224-by-224 |
Xception [88] |
“xception” |
71 |
85 |
22.9 |
299-by-299 |
Table 5.
Parameters of network performance of the test dataset (holdout, new data).
Table 5.
Parameters of network performance of the test dataset (holdout, new data).
Predicted variable |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
Specificity (%) |
False Positive Rate (%) |
Ulcerative colitis |
99.10 |
97.09 |
94.79 |
95.93 |
99.64 |
0.36 |
Adenocarcinoma |
99.84 |
99.90 |
99.88 |
99.89 |
99.70 |
0.30 |
Colon control |
99.06 |
95.75 |
97.63 |
96.68 |
99.29 |
0.71 |
Table 6.
Parameters of network performance of the test dataset (holdout, new data) using H&E images.
Table 6.
Parameters of network performance of the test dataset (holdout, new data) using H&E images.
Predicted variable |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
Specificity (%) |
False Positive Rate (%) |
Steroid-requiring |
79.53 |
66.18 |
70.74 |
68.39 |
83.53 |
16.47 |
Mesalazine-responsive |
79.53 |
86.23 |
83.53 |
84.86 |
70.74 |
29.26 |
Table 7.
Parameters of network performance of the test dataset (holdout, new data) using LAIR1 immunohistochemistry.
Table 7.
Parameters of network performance of the test dataset (holdout, new data) using LAIR1 immunohistochemistry.
Predicted variable |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
Specificity (%) |
False Positive Rate (%) |
Steroid-requiring |
88.31 |
79.38 |
82.26 |
80.79 |
90.89 |
9.11 |
Mesalazine-responsive |
88.31 |
92.31 |
90.89 |
91.59 |
82.26 |
17.74 |
Table 8.
Parameters of network performance of the test dataset (holdout, new data) using TOX2 immunohistochemistry.
Table 8.
Parameters of network performance of the test dataset (holdout, new data) using TOX2 immunohistochemistry.
Predicted variable |
Accuracy (%) |
Precision (%) |
Recall (%) |
F1-Score (%) |
Specificity (%) |
False Positive Rate (%) |
Steroid-requiring |
85.62 |
72.04 |
79.51 |
75.59 |
87.99 |
12.01 |
Mesalazine-responsive |
85.62 |
91.69 |
87.99 |
89.80 |
79.51 |
20.49 |
Table 9.
Performance comparison of CNN networks on the test set (holdout, new data) using H&E images.
Table 9.
Performance comparison of CNN networks on the test set (holdout, new data) using H&E images.
Model |
Accuracy (%) |
DenseNet-201 |
99.30 |
ResNet-50 |
99.14 |
Inception-v3 |
99.13 |
ResNet-101 |
99.10 |
ResNet-18 |
99.00 |
ShuffleNet |
98.94 |
MobileNet-v2 |
98.89 |
NasNet-Large |
98.88 |
GoogLeNet-Places365 |
98.86 |
VGG-19 |
98.80 |
EfficientNet-b0 |
98.79 |
AlexNet |
98.77 |
Xception |
98.66 |
VGG-16 |
98.65 |
GoogLeNet |
98.60 |
NasNet-Mobile |
98.58 |