Capsule endoscopy has been widely used as a non-invasive diagnostic tool for small or large intestinal lesions. In recent years, automated lesion detection systems using machine learning have been devised. This study aimed to develop an automated system for capsule endoscopic severity in patients with ulcerative colitis along the entire length of the colon using ResNet50. Capsule endoscopy videos from patients with ulcerative colitis were collected prospectively. Each single examination video file was partitioned into four segments: the cecum and ascending colon, transverse colon, descending and sigmoid colon, and rectum. Fifty still pictures (576 × 576 pixels) were extracted from each partitioned video. A patch (128 × 128 pixels) was trimmed from the still picture at every 32-pixel-strides. A total of 739,021 patch images were manually classified into six categories: 0) Mayo endoscopic subscore (MES) 0, 1) MES1, 2) MES2, 3) MES3, 4) inadequate quality for evaluation, and 5) ileal mucosa. ResNet50, a deep learning framework, was trained using 483,644 datasets and validated using 255,377 independent datasets. In total, 31 capsule endoscopy videos from 22 patients were collected. The accuracy rates of the training and validation datasets were 0.992 and 0.973, respectively. An automated evaluation system for the capsule endoscopic severity of ulcerative colitis was developed. This could be a useful tool for assessing topographic disease activity, thus decreasing the burden of image interpretation on endoscopists.

Massive intraoperative blood loss (IBL) negatively influence outcomes after surgery for pancreatic ductal adenocarcinoma (PDAC). However, few data or predictive models are available for the identification of patients with a high risk for massive IBL. This study aimed to build a model for massive IBL prediction using a decision tree algorithm, which is one machine learning method. One hundred and seventy-five patients undergoing curative surgery for resectable PDAC at our facility between January 2007 and October 2020 were allocated to training (n = 128) and testing (n = 47) sets. Using the preoperatively available data of the patients (34 variables), we built a decision tree classification algorithm. Of the 175 patients, massive IBL occurred in 88 patients (50.3%). Binary logistic regression analysis indicated that alanine aminotransferase and distal pancreatectomy were significant predictors of massive IBL occurrence with an overall correct prediction rate of 70.3%. Decision tree analysis automatically selected 14 predictive variables. The best predictor was the surgical procedure. Though massive IBL was not common, the outcome of patients with distal pancreatectomy was secondarily split by glutamyl transpeptidase. Among patients who underwent PD (n = 83), diabetes mellitus (DM) was selected as the variable in the second split. Of the 21 patients with DM, massive IBL occurred in 85.7%. Decision tree sensitivity was 98.5% in the training data set and 100% in the testing data set. Our findings suggested that a decision tree can provide a new potential approach to predict massive IBL in surgery for resectable PDAC.

Surgery avoidance is an important goal in Crohn’s disease (CD) treatment and predicting the risk of subsequent surgery is important to determine adequate therapeutic strength for patients with newly diagnosed CD. Herein, we aimed to construct a prediction model for the risk of subsequent surgery based on disease characteristics at the patients’ initial visit. We retrospectively collected disease characteristic data from 93 patients with newly diagnosed CD. A logistic regression model with a brute force method was used to maximize the area under the receiver operating characteristic curve (auROC) by employing a combination of potential predictors from 14 covariates (16,383). The auROC remained almost constant when one to 12 covariates were considered, reaching a peak of 0.89 at four covariates (small-bowel patency, extensive small-bowel lesions, main lesions, and the number of poor prognostic factors), and it decreased with increasing covariate size. The most significant predictors were small-bowel patency, extensive small-bowel lesions, and age or major lesions. Therefore, this prediction model using covariates may be helpful in determining the likelihood that a patient with newly diagnosed CD will require surgery, which can aid in appropriate treatment selection for high-risk patients.

Narrow band imaging (NBI) has been extensively utilized as a diagnostic tool for colorectal neoplastic lesions. This study aimed to develop a trial deep learning (DL) based four-class classification model for low-grade dysplasia (LGD); high-grade dysplasia or mucosal carcinoma (HGD); superficially invasive submucosal carcinoma (SMs) and deeply invasive submucosal carcinomas (SMd) and evaluate its potential as a diagnostic tool. We collected a total of 1,390 NBI images as the dataset, including 53 LGD, 120 HGD, 20 SMs and 17 SMd. A total of 598,801 patches were trimmed from the lesion and background. A patch-based classification model was built by employing a residual convolutional neural network (CNN) and validated by three-fold cross-validation. The patch-based validation accuracy was 0.876, 0.957, 0.907 and 0.929 in LGD, HGD, SMs and SMd, respectively. The image-level classification algorithm was derived from the patch-based mapping across the entire image domain, attaining accuracies of 0.983, 0.990, 0.964, and 0.992 in LGD, HGD, SMs, and SMd, respectively. Our CNN-based model demonstrated high performance for categorizing the histological grade of dysplasia as well as the depth of invasion in routine colonoscopy, suggesting a potential diagnostic tool with minimal human inputs.

Narrow band imaging (NBI) has been extensively utilized as a diagnostic tool for colorectal neoplastic lesions. This study aimed to develop a trial deep learning (DL) based four-class classification model for low-grade dysplasia (LGD); high-grade dysplasia or mucosal carcinoma (HGD); superficially invasive submucosal carcinoma (SMs) and deeply invasive submucosal carcinomas (SMd) and evaluate its potential as a diagnostic tool. We collected a total of 1390 NBI images as the dataset, including 53 LGD, 120 HGD, 20 SMs and 17 SMd. A total of 59,8801 patches were trimmed from the lesion and background. A patch-based classification model was built by employing a residual convolutional neural network (CNN) and validated by three-fold cross-validation. The patch-based validation accuracy was 0.876, 0.957, 0.907 and 0.929 in LGD, HGD, SMs and SMd, respectively. The image-level classification algorithm was derived from the patch-based mapping across the entire image domain, attaining accuracies of 0.983, 0.990, 0.964, and 0.992 in LGD, HGD, SMs, and SMd, respectively. Our CNN-based model demonstrated high performance for categorizing the histological grade of dysplasia as well as the depth of invasion in routine colonoscopy, suggesting a potential diagnostic tool with minimal human inputs.