Automated Classification of Colorectal Neoplasms in White-Light Colonoscopy Images via Deep Learning

Background: Classification of colorectal neoplasms during colonoscopic examination is important to avoid unnecessary endoscopic biopsy or resection. This study aimed to develop and validate deep learning models that automatically classify colorectal lesions histologically on white-light colonoscopy images. Methods: White-light colonoscopy images of colorectal lesions exhibiting pathological results were collected and classified into seven categories: stages T1-4 colorectal cancer (CRC), high-grade dysplasia (HGD), tubular adenoma (TA), and non-neoplasms. The images were then re-classified into four categories including advanced CRC, early CRC/HGD, TA, and non-neoplasms. Two convolutional neural network models were trained, and the performances were evaluated in an internal test dataset and an external validation dataset. Results: In total, 3828 images were collected from 1339 patients. The mean accuracies of ResNet-152 model for the seven-category and four-category classification were 60.2% and 67.3% in the internal test dataset, and 74.7% and 79.2% in the external validation dataset, respectively, including 240 images. In the external validation, ResNet-152 outperformed two endoscopists for four-category classification, and showed a higher mean area under the curve (AUC) for detecting TA+ lesions (0.818) compared to the worst-performing endoscopist. The mean AUC for detecting HGD+ lesions reached 0.876 by Inception-ResNet-v2. Conclusions: A deep learning model presented promising performance in classifying colorectal lesions on white-light colonoscopy images; this model could help endoscopists build optimal treatment strategies.

Download Full-text

Deep learning based prediction of prognosis in nonmetastatic clear cell renal cell carcinoma

Scientific Reports ◽

10.1038/s41598-020-80262-9 ◽

2021 ◽

Vol 11 (1) ◽

Author(s):

Seok-Soo Byun ◽

Tak Sung Heo ◽

Jeong Myeong Choi ◽

Yeong Seok Jeong ◽

Yu Seop Kim ◽

...

Keyword(s):

Renal Cell Carcinoma ◽

Deep Learning ◽

Cell Carcinoma ◽

Renal Cell ◽

Clear Cell ◽

External Validation ◽

Training Dataset ◽

Validation Dataset ◽

Cell Renal Cell Carcinoma ◽

Test Dataset

AbstractSurvival analyses for malignancies, including renal cell carcinoma (RCC), have primarily been conducted using the Cox proportional hazards (CPH) model. We compared the random survival forest (RSF) and DeepSurv models with the CPH model to predict recurrence-free survival (RFS) and cancer-specific survival (CSS) in non-metastatic clear cell RCC (nm-cRCC) patients. Our cohort included 2139 nm-cRCC patients who underwent curative-intent surgery at six Korean institutions between 2000 and 2014. The data of two largest hospitals’ patients were assigned into the training and validation dataset, and the data of the remaining hospitals were assigned into the external validation dataset. The performance of the RSF and DeepSurv models was compared with that of CPH using Harrel’s C-index. During the follow-up, recurrence and cancer-specific deaths were recorded in 190 (12.7%) and 108 (7.0%) patients, respectively, in the training-dataset. Harrel’s C-indices for RFS in the test-dataset were 0.794, 0.789, and 0.802 for CPH, RSF, and DeepSurv, respectively. Harrel’s C-indices for CSS in the test-dataset were 0.831, 0.790, and 0.834 for CPH, RSF, and DeepSurv, respectively. In predicting RFS and CSS in nm-cRCC patients, the performance of DeepSurv was superior to that of CPH and RSF. In no distant time, deep learning-based survival predictions may be useful in RCC patients.

Download Full-text

Deep Learning for Caries Detection and Classification

Diagnostics ◽

10.3390/diagnostics11091672 ◽

2021 ◽

Vol 11 (9) ◽

pp. 1672

Author(s):

Luya Lian ◽

Tianer Zhu ◽

Fudong Zhu ◽

Haihua Zhu

Keyword(s):

Neural Network ◽

Deep Learning ◽

Disease Diagnosis ◽

Validation Dataset ◽

Reference Dataset ◽

Dice Coefficient ◽

Learning Methods ◽

Test Dataset ◽

The Impact ◽

Caries Lesions

Objectives: Deep learning methods have achieved impressive diagnostic performance in the field of radiology. The current study aimed to use deep learning methods to detect caries lesions, classify different radiographic extensions on panoramic films, and compare the classification results with those of expert dentists. Methods: A total of 1160 dental panoramic films were evaluated by three expert dentists. All caries lesions in the films were marked with circles, whose combination was defined as the reference dataset. A training and validation dataset (1071) and a test dataset (89) were then established from the reference dataset. A convolutional neural network, called nnU-Net, was applied to detect caries lesions, and DenseNet121 was applied to classify the lesions according to their depths (dentin lesions in the outer, middle, or inner third D1/2/3 of dentin). The performance of the test dataset in the trained nnU-Net and DenseNet121 models was compared with the results of six expert dentists in terms of the intersection over union (IoU), Dice coefficient, accuracy, precision, recall, negative predictive value (NPV), and F1-score metrics. Results: nnU-Net yielded caries lesion segmentation IoU and Dice coefficient values of 0.785 and 0.663, respectively, and the accuracy and recall rate of nnU-Net were 0.986 and 0.821, respectively. The results of the expert dentists and the neural network were shown to be no different in terms of accuracy, precision, recall, NPV, and F1-score. For caries depth classification, DenseNet121 showed an overall accuracy of 0.957 for D1 lesions, 0.832 for D2 lesions, and 0.863 for D3 lesions. The recall results of the D1/D2/D3 lesions were 0.765, 0.652, and 0.918, respectively. All metric values, including accuracy, precision, recall, NPV, and F1-score values, were proven to be no different from those of the experienced dentists. Conclusion: In detecting and classifying caries lesions on dental panoramic radiographs, the performance of deep learning methods was similar to that of expert dentists. The impact of applying these well-trained neural networks for disease diagnosis and treatment decision making should be explored.

Download Full-text

Comparison of deep learning and conventional machine learning methods for classification of colon polyp types

The EuroBiotech Journal ◽

10.2478/ebtj-2021-0006 ◽

2021 ◽

Vol 5 (1) ◽

pp. 34-42

Author(s):

Refika Sultan Doğan ◽

Bülent Yılmaz

Keyword(s):

Machine Learning ◽

Deep Learning ◽

White Light ◽

Narrow Band ◽

Optical Biopsy ◽

Colon Polyps ◽

Tissue Biopsy ◽

Conventional Machine ◽

White Light Imaging

AbstractDetermination of polyp types requires tissue biopsy during colonoscopy and then histopathological examination of the microscopic images which tremendously time-consuming and costly. The first aim of this study was to design a computer-aided diagnosis system to classify polyp types using colonoscopy images (optical biopsy) without the need for tissue biopsy. For this purpose, two different approaches were designed based on conventional machine learning (ML) and deep learning. Firstly, classification was performed using random forest approach by means of the features obtained from the histogram of gradients descriptor. Secondly, simple convolutional neural networks (CNN) based architecture was built to train with the colonoscopy images containing colon polyps. The performances of these approaches on two (adenoma & serrated vs. hyperplastic) or three (adenoma vs. hyperplastic vs. serrated) category classifications were investigated. Furthermore, the effect of imaging modality on the classification was also examined using white-light and narrow band imaging systems. The performance of these approaches was compared with the results obtained by 3 novice and 4 expert doctors. Two-category classification results showed that conventional ML approach achieved significantly better than the simple CNN based approach did in both narrow band and white-light imaging modalities. The accuracy reached almost 95% for white-light imaging. This performance surpassed the correct classification rate of all 7 doctors. Additionally, the second task (three-category) results indicated that the simple CNN architecture outperformed both conventional ML based approaches and the doctors. This study shows the feasibility of using conventional machine learning or deep learning based approaches in automatic classification of colon types on colonoscopy images.

Download Full-text

Deep Learning Model for Screening Sepsis Using Electrocardiography

10.21203/rs.3.rs-186976/v1 ◽

2021 ◽

Author(s):

Joon-myoung Kwon ◽

Ye Rang Lee ◽

Min-Seung Jung ◽

Yoon-Ji Lee ◽

Yong-Yeon Jo ◽

...

Keyword(s):

Septic Shock ◽

Deep Learning ◽

Characteristic Curve ◽

External Validation ◽

Medical Emergency ◽

Validation Dataset ◽

Internal Validation ◽

Significant Difference ◽

Life Threatening ◽

Sepsis And Septic Shock

Abstract Background: Sepsis is a life-threatening organ dysfunction and is a major healthcare burden worldwide. Although sepsis is a medical emergency that requires immediate management, it is difficult to screen the occurrence of sepsis. In this study, we propose an artificial intelligence based on deep learning-based model (DLM) for screening sepsis using electrocardiography (ECG).Methods: This retrospective cohort study included 46,017 patients who admitted to two hospitals. 1,548 and 639 patients underwent sepsis and septic shock. The DLM was developed using 73,727 ECGs of 18,142 patients and internal validation was conducted using 7,774 ECGs of 7,774 patients. Furthermore, we conducted an external validation with 20,101 ECGs of 20,101 patients from another hospital to verify the applicability of the DLM across centers.Results: During the internal and external validation, the area under the receiver operating characteristic curve (AUC) of an DLM using 12-lead ECG for screening sepsis were 0.901 (95% confidence interval 0.882–0.920) and 0.863 (0.846–0.879), respectively. During internal and external validation, AUC of an DLM for detecting septic shock were 0.906 (95% CI = 0.877–0.936) and 0.899 (95% CI = 0.872–0.925), respectively. The AUC of the DLM for detecting sepsis using 6-lead and single-lead ECGs were 0.845–0.882. A sensitivity map showed that the QRS complex and T wave was associated with sepsis. Subgroup analysis was conducted using ECGs from 4,609 patients who admitted with infectious disease, The AUC of the DLM for predicting in-hospital mortality was 0.817 (0.793–0.840). There was a significant difference in the prediction score of DLM using ECG according to the presence of infection in the validation dataset (0.277 vs 0.574, p<0.001), including severe acute respiratory syndrome coronavirus 2 (0.260 vs 0.725, p=0.018).Conclusions: The DLM demonstrated reasonable performance for screening sepsis using 12-, 6-, and single-lead ECG. The results suggest that sepsis can be screened using not only conventional ECG devices, but also diverse life-type ECG machine employing the DLM, thereby preventing irreversible disease progression and mortality.

Download Full-text

A Rapid Motor Task-Based Screening Tool for Parkinsonism in Community-Based Studies

Frontiers in Neurology ◽

10.3389/fneur.2021.653066 ◽

2021 ◽

Vol 12 ◽

Author(s):

Wendy W. Dlamini ◽

Searles Nielsen ◽

Mwiza Ushe ◽

Gill Nelson ◽

Brad A. Racette

Keyword(s):

Standard Deviation ◽

Predictive Model ◽

Screening Tool ◽

Motor Task ◽

Population Based ◽

Validation Dataset ◽

Motor Tasks ◽

Test Dataset ◽

The Mean ◽

Summary Measures

Background: The prevalence of parkinsonism in developing countries is largely unknown due to difficulty in ascertainment because access to neurologists is often limited.Objective: Develop and validate a parkinsonism screening tool using objective motor task-based tests that can be administered by non-clinicians.Methods: In a cross-sectional population-based sample from South Africa, we evaluated 315 adults, age >40, from an Mn-exposed (smelter) community, using the Unified Parkinson Disease Rating Scale motor subsection 3 (UPDRS3), Purdue grooved pegboard, and kinematic-UPDRS3-based motor tasks. In 275 participants (training dataset), we constructed a linear regression model to predict UPDRS3. We selected motor task summary measures independently associated with UPDRS3 (p < 0.05). We validated the model internally in the remaining 40 participants from the manganese-exposed community (test dataset) using the area under the receiver operating characteristic curve (AUC), and externally in another population-based sample of 90 participants from another South African community with only background levels of environmental Mn exposure.Results: The mean UPDRS3 score in participants from the Mn-exposed community was 9.1 in both the training and test datasets (standard deviation = 6.4 and 6.1, respectively). Together, 57 (18.1%) participants in this community had a UPDRS3 ≥ 15, including three with Parkinson's disease. In the non-exposed community, the mean UPDRS3 was 3.9 (standard deviation = 4.3). Three (3.3%) had a UPDRS3 ≥ 15. Grooved pegboard time and mean velocity for hand rotation and finger tapping tasks were strongly associated with UPDRS3. Using these motor task summary measures and age, the UPDRS3 predictive model performed very well. In the test dataset, AUCs were 0.81 (95% CI 0.68, 0.94) and 0.91 (95% CI 0.81, 1.00) for cut points for neurologist-assessed UPDRS3 ≥ 10 and UPDRS3 ≥ 15, respectively. In the external validation dataset, the AUC was 0.85 (95% CI 0.73, 0.97) for UPDRS3 ≥ 10. AUCs were 0.76–0.82 when excluding age.Conclusion: A predictive model based on a series of objective motor tasks performs very well in assessing severity of parkinsonism in both Mn-exposed and non-exposed population-based cohorts.

Download Full-text

Automated classification of gastric neoplasms in endoscopic images using a convolutional neural network

Endoscopy ◽

10.1055/a-0981-6133 ◽

2019 ◽

Vol 51 (12) ◽

pp. 1121-1129 ◽

Cited By ~ 17

Author(s):

Bum-Joo Cho ◽

Chang Seok Bang ◽

Se Woo Park ◽

Young Joo Yang ◽

Seung In Seo ◽

...

Keyword(s):

Neural Network ◽

Gastric Cancer ◽

Convolutional Neural Network ◽

White Light ◽

Lesion Detection ◽

Gastric Neoplasms ◽

Validation Dataset ◽

Low Grade ◽

Test Dataset ◽

Endoscopic Images

Abstract Background Visual inspection, lesion detection, and differentiation between malignant and benign features are key aspects of an endoscopist’s role. The use of machine learning for the recognition and differentiation of images has been increasingly adopted in clinical practice. This study aimed to establish convolutional neural network (CNN) models to automatically classify gastric neoplasms based on endoscopic images. Methods Endoscopic white-light images of pathologically confirmed gastric lesions were collected and classified into five categories: advanced gastric cancer, early gastric cancer, high grade dysplasia, low grade dysplasia, and non-neoplasm. Three pretrained CNN models were fine-tuned using a training dataset. The classifying performance of the models was evaluated using a test dataset and a prospective validation dataset. Results A total of 5017 images were collected from 1269 patients, among which 812 images from 212 patients were used as the test dataset. An additional 200 images from 200 patients were collected and used for prospective validation. For the five-category classification, the weighted average accuracy of the Inception-Resnet-v2 model reached 84.6 %. The mean area under the curve (AUC) of the model for differentiating gastric cancer and neoplasm was 0.877 and 0.927, respectively. In prospective validation, the Inception-Resnet-v2 model showed lower performance compared with the endoscopist with the best performance (five-category accuracy 76.4 % vs. 87.6 %; cancer 76.0 % vs. 97.5 %; neoplasm 73.5 % vs. 96.5 %; P < 0.001). However, there was no statistical difference between the Inception-Resnet-v2 model and the endoscopist with the worst performance in the differentiation of gastric cancer (accuracy 76.0 % vs. 82.0 %) and neoplasm (AUC 0.776 vs. 0.865). Conclusion The evaluated deep-learning models have the potential for clinical application in classifying gastric cancer or neoplasm on endoscopic white-light images.

Download Full-text

Development and validation of a deep learning algorithm detecting 10 common abnormalities on chest radiographs

European Respiratory Journal ◽

10.1183/13993003.03061-2020 ◽

2020 ◽

pp. 2003061

Author(s):

Ju Gang Nam ◽

Minchul Kim ◽

Jongchan Park ◽

Eui Jin Hwang ◽

Jong Hyuk Lee ◽

...

Keyword(s):

Emergency Department ◽

Deep Learning ◽

Learning Algorithm ◽

External Validation ◽

Chest Radiographs ◽

Reading Tests ◽

Deep Learning Algorithm ◽

The Mean ◽

Emergency Department Patients ◽

Development And Validation

We aimed to develop a deep-learning algorithm detecting 10 common abnormalities (DLAD-10) on chest radiographs and to evaluate its impact in diagnostic accuracy, timeliness of reporting, and workflow efficacy.DLAD-10 was trained with 146 717 radiographs from 108 053 patients using a ResNet34-based neural network with lesion-specific channels for 10 common radiologic abnormalities (pneumothorax, mediastinal widening, pneumoperitoneum, nodule/mass, consolidation, pleural effusion, linear atelectasis, fibrosis, calcification, and cardiomegaly). For external validation, the performance of DLAD-10 on a same-day CT-confirmed dataset (normal:abnormal, 53:147) and an open-source dataset (PadChest; normal:abnormal, 339:334) was compared to that of three radiologists. Separate simulated reading tests were conducted on another dataset adjusted to real-world disease prevalence in the emergency department, consisting of four critical, 52 urgent, and 146 non-urgent cases. Six radiologists participated in the simulated reading sessions with and without DLAD-10.DLAD-10 exhibited areas under the receiver-operating characteristic curves (AUROCs) of 0.895–1.00 in the CT-confirmed dataset and 0.913–0.997 in the PadChest dataset. DLAD-10 correctly classified significantly more critical abnormalities (95.0% [57/60]) than pooled radiologists (84.4% [152/180]; p=0.01). In simulated reading tests for emergency department patients, pooled readers detected significantly more critical (70.8% [17/24] versus 29.2% [7/24]; p=0.006) and urgent (82.7% [258/312] versus 78.2% [244/312]; p=0.04) abnormalities when aided by DLAD-10. DLAD-10 assistance shortened the mean time-to-report critical and urgent radiographs (640.5±466.3 versus 3371.0±1352.5 s and 1840.3±1141.1 versus 2127.1±1468.2, respectively; p-values<0.01) and reduced the mean interpretation time (20.5±22.8 versus 23.5±23.7 s; p<0.001).DLAD-10 showed excellent performance, improving radiologists' performance and shortening the reporting time for critical and urgent cases.

Download Full-text

Deep Learning Analysis to Automatically Detect the Presence of Penetration or Aspiration in Videofluoroscopic Swallowing Study

10.21203/rs.3.rs-884186/v1 ◽

2021 ◽

Author(s):

Jeoung Kun Kim ◽

Yoo Jin Choo ◽

Gyu Sang Choi ◽

Hyunkwang Shin ◽

Min Cheol Chang ◽

...

Keyword(s):

Deep Learning ◽

Gold Standard ◽

Area Under The Curve ◽

High Peak ◽

Validation Dataset ◽

Videofluoroscopic Swallowing Study ◽

Python Programming Language ◽

Python Programming ◽

Learning Analysis

Abstract Background: Videofluoroscopic swallowing study (VFSS) is currently considered the gold standard to precisely diagnose and quantitatively investigate dysphagia. However, VFSS interpretation is complex and requires consideration of several factors. Purpose: Therefore, considering the expected impact on dysphagia management, this study aimed to apply deep learning to detect the presence of penetration or aspiration in VFSS of patients with dysphagia automatically.Materials and Methods: The VFSS data of 190 participants with dysphagia were collected. A total of 10 frame images from one swallowing process were selected (five high-peak images and five low-peak images) for the application of deep learning in a VFSS video of a patient with dysphagia. We applied a convolutional neural network (CNN) for deep learning using the Python programming language. For the classification of VFSS findings (normal swallowing, penetration, and aspiration), the classification was determined in both high-peak and low-peak images. Thereafter, the two classifications determined through high-peak and low-peak images were integrated into a final classification.Results: The area under the curve (AUC) for the validation dataset of the VFSS image for the CNN model was 0.946 for normal findings, 0.885 for penetration, and 1.000 for aspiration. The average AUC was 0.962.Conclusion: This study demonstrated that deep learning algorithms, particularly the CNN, could be applied for detecting the presence of penetration and aspiration in VFSS of patients with dysphagia.

Download Full-text

Application of Deep Learning for 3D building generalization

Proceedings of the ICA ◽

10.5194/ica-proc-2-147-2019 ◽

2019 ◽

Vol 2 ◽

pp. 1-8

Author(s):

Yue Wu ◽

Yevgeniya Filippovska ◽

Valentina Schmidt ◽

Martin Kada

Keyword(s):

Deep Learning ◽

Loss Function ◽

Semantic Content ◽

Classification Problems ◽

Optimal Model ◽

Test Dataset ◽

And Topology ◽

Generalization Problem ◽

Ablation Study

<p><strong>Abstract.</strong> The generalization of 3D buildings is a challenging task, which needs to consider geometry information, semantic content and topology relations of 3D buildings. Although many algorithms with detailed and reasonable designs have been developed for the 3D building generalization, there are still cases that could be further studied. As a fast-growing technique, Deep Learning has shown its ability to build complex concepts out of simpler concepts in many fields. Therefore, in this paper, Deep Learning is used to solve the regression (generalization of individual 3D building) and classification problems (classification of roof type) simultaneously. Firstly, the test dataset is generated and labelled with the generalization results as well as the classification of roof types. Buildings with saddleback, half-hip, and hip roof are selected as the experimental objects since their generalization results can be uniformly represented by a common vector which aims to meet the compatible representation of Deep Learning. Then, the pre-trained ResNet50 is used as the baseline network. The optimal model capacity is searched within an extensive ablation study in the framework of the building generalization problem. After that, a multi-task network is built by adding a branch of classification to the above network, which is in parallel with the generalization branch. In the process of training, the imbalance problems of tasks and classes are solved by adjusting their donations to the total loss function. It is found that less error rate is obtained after adding a classification branch. For the final results, two improved metrics are used to evaluate the generalization performance. The accuracy of generalization reached over 95% for horizontal information and 85% for height, while the accuracy of classification reached 100% on the test dataset.</p>

Download Full-text

Predicting Sex from Retinal Fundus Photographs Using Automated Deep Learning

10.21203/rs.3.rs-402433/v1 ◽

2021 ◽

Author(s):

Edward Korot ◽

Nikolas Pontikos ◽

Xiaoxuan Liu ◽

Siegfried K Wagner ◽

Livia Faes ◽

...

Keyword(s):

Deep Learning ◽

Characteristic Curve ◽

External Validation ◽

Model Development ◽

Model Performance ◽

Validation Dataset ◽

Internal Validation ◽

Fundus Photographs ◽

The Uk ◽

Retinal Fundus

Abstract Deep learning may transform health care, but model development has largely been dependent on availability of advanced technical expertise. Herein we present the development of a deep learning model by clinicians without coding, which predicts reported sex from retinal fundus photographs. A model was trained on 84,743 retinal fundus photos from the UK Biobank dataset. External validation was performed on 252 fundus photos from a tertiary ophthalmic referral center. For internal validation, the area under the receiver operating characteristic curve (AUROC) of the code free deep learning (CFDL) model was 0.93. Sensitivity, specificity, positive predictive value (PPV) and accuracy (ACC) were 88.8%, 83.6%, 87.3% and 86.5%, and for external validation were 83.9%, 72.2%, 78.2% and 78.6% respectively. Clinicians are currently unaware of distinct retinal feature variations between males and females, highlighting the importance of model explainability for this task. The model performed significantly worse when foveal pathology was present in the external validation dataset, ACC: 69.4%, compared to 85.4% in healthy eyes, suggesting the fovea is a salient region for model performance OR (95% CI): 0.36 (0.19, 0.70) p = 0.0022. Automated machine learning (AutoML) may enable clinician-driven automated discovery of novel insights and disease biomarkers.

Download Full-text