Ensemble Learning Approach Research Articles

Skin Disease Detection Using Ensemble Learning

Skin diseases affect millions of people worldwide, making early detection and accurate diagnosis crucial for effective treatment. However, the process of diagnosing skin conditions often requires specialized dermatological expertise, which can be time-consuming and expensive. To address this challenge, we have developed a web-based Skin Disease Detection System using cutting-edge machine learning and ensemble techniques. Our system leverages the HAM10000 dataset, consisting of 10,000 dermoscopic images representing seven common skin lesion types. We employ an ensemble learning approach, integrating ResNet50, EfficientNetB0, and DenseNet121 as base models and combining their predictions using a logistic regression meta-model. This architecture enhances the accuracy and robustness of the predictions. The backend of our application is built with Flask, responsible for model processing, authentication, and handling API requests. The React frontend provides an intuitive user interface where users can upload skin lesion images and receive real-time predictions along with confidence scores. The system also incorporates authentication and token verification for secure user access, ensuring data privacy and integrity. Our project aims to provide an accessible, reliable, and cost-effective tool to assist healthcare professionals and individuals in the early detection of skin diseases, potentially improving patient outcomes and reducing the burden on dermatology services

Exploring Thematic Diversity in Classical Chinese Poetry: A Novel Dataset and a BERT-Enhanced Ensemble Learning Approach

Classical Chinese poetry, as an essential aspect of cultural heritage, exhibits rich theme diversity often overlooked in natural language processing research. To address this gap, we aim to explore the classification of thematic categories within this literary domain. We curate a dataset of 2,918 annotated poems spanning 7 common themes and propose a BERT-based ensemble learning approach for effective classification. Although this method integrates existing models, it achieves an accuracy and F1 score of over 72% in the 7-class task, surpassing established baselines, and providing a baseline for future research. The experimental findings reveal the effectiveness of ensemble strategies in improving individual base model performance and highlight the potential of the MLP-based ensemble technique. The study contributes to a deeper understanding of thematic categories and textual features in classical Chinese poetry and offers an automated classification system for classical Chinese poems.

Heterogeneous knowledge graph-driven subassembly identification with ensemble deep learning in Industry 4.0

In the context of Industry 4.0, model-based definition (MBD) has been an effective approach to creating 3D models contained all heterogeneous information needed to define a product, which proposes new challenges for the traditional subassembly identification method that only considers the geometric information of a product in assembly sequence planning. To bridge the gap, we propose a novel heterogeneous knowledge graph-driven subassembly identification method to enhance assembly sequence planning in the model-based systems engineering (MBSE) paradigm. Specifically, a heterogeneous knowledge graph is first constructed based on the shape information and engineering details of an MBD model. Next, an ensemble deep learning approach that combines graph neural networks with the community detection algorithm is proposed to effectively detect the subassembly from the MBD model. Finally, the feasibility and effectiveness of the proposed method are demonstrated through an example of car suspension subassembly identification, providing insight into the industrial implementation.

DETECTION OF DUST ON SOLAR PANELS WITH DEEP LEARNING

Solar energy is an environmentally friendly, clean, and sustainable alternative. The widespread use of this energy source offers excellent environmental and economic benefits. However, some factors affect the efficiency of solar panels. One of these factors is dust. When dust accumulates on the surface of solar panels, it can significantly reduce the efficiency of energy production. Therefore, detecting and quickly removing dust from solar panels is crucial. Managing this process with unmanned artificial intelligence systems, especially in large areas, will provide significant advantages in terms of time and cost. In recent years, convolutional neural networks have achieved significant success in image classification. In particular, transfer learning methods have proven their success in this field. In this study, we aim to solve a new task with limited data using pre-trained deep learning models (EfficientNetB3, ResNet50, MobileNet, VGG19, Xception, InceptionResNetV2, VGG16, ResNet101, DenseNet201, EfficientNetB7) to classify dirty and clean solar panels. These models were chosen because they each have different strengths and have performed well on various tasks. The models with the best performance among these models are combined to improve classification prediction. The proposed ensemble learning approach achieved 99.31% classification accuracy by considering the prediction results of the models with a voting approach. As a result, this approach aims to optimize the maintenance processes of solar energy systems, improve energy efficiency, and support sustainable energy use in the long term.

Stacking-based ensemble learning for identifying artist signatures on paintings

Identifying artist signatures on paintings is essential for authenticating artworks and advancing digital humanities. An artist’s signature is a consistent element included in each painting that the artist creates, providing a unique identifier for their work. Traditional methods that rely on expert analysis and manual comparison are time-consuming and are prone to human error. Although convolutional neural networks (CNNs) have shown promise in automating this process, existing single-model approaches struggle with the diversity and complexity of artistic styles, leading to limitations in their performance and generalizability. Therefore, this study proposes an ensemble learning approach that integrates the predictive power of multiple CNN-based models. The proposed framework leverages the strengths of three state-of-the-art CNNs: EfficientNetB4, ResNet-50, and Xception. These models were independently trained, and the predictions were combined using a meta-learning strategy. To address class imbalance, data augmentation techniques and weighted loss functions were employed. The experimental results obtained on a dataset of more than 8,000 paintings from 50 artists demonstrate significant improvements over individual CNN architectures and other ensemble methods, thereby effectively capturing complex features and improving generalizability.

Improving signal-to-noise ratio of Raman measurements based on ensemble learning approach.

Raman spectroscopy is an extensively explored vibrational spectroscopic technique to analyzethebiochemical composition and molecular structure of samples, which is often assumed to be non-destructive when carefully using proper laser power and exposure time. However, the inherentlyweak Raman signal and concurrent fluorescence interference often lead to Raman measurements withalow signal-to-noise ratio (SNR), especially for biological samples. Great efforts have been made to develop experimental approaches and/or numerical algorithms to improvetheSNR. In this study, we proposed an ensemble learning approach to recover and denoise Raman measurements withalow SNR. The proposed ensemble learning approach was evaluated on 986 pairs of Raman measurements, each pair of which consists of a low SNR Raman spectrum and a high SNR reference Raman spectrum from the exact same fungal sample but uses 200 timestheintegration time. Compared with conventional methods, the Raman measurementsrecoveredby the proposed ensemble learning approach are more identical to high SNR reference Raman measurements, with anaverage RMSE and MAE of only 1.337 × 10-2 and 1.066 × 10-2, respectively; thus, the proposed ensemble learning approach is expected to be a powerful tool for numerically improving the SNR of Raman measurements and further benefits rapid Raman acquisition from biological samples.

Enhancing Cardiovascular Disease Prediction Accuracy through an Ensemble Machine Learning Approach

This study explores the efficacy of an ensemble machine learning approach, specifically a Voting Classifier combining Decision Tree, k-Nearest Neighbors, and Gaussian Naive Bayes classifiers, in predicting cardiovascular diseases (CVDs). Utilizing a dataset consisting of 70,000 clinical records, the model was rigorously tested through 5-fold cross-validation, achieving remarkable results with average accuracies, precision, recall, and F1-scores all exceeding 99%. The findings validate the hypothesis that ensemble models, due to their capacity to leverage multiple learning algorithms, provide superior prediction accuracy and reliability compared to single predictor models. This research not only confirms the effectiveness of ensemble methods in medical diagnostics but also highlights their potential to enhance decision-making in clinical settings. Given the model's success in identifying various stages of cardiovascular conditions with high accuracy, it offers significant implications for early intervention and personalized patient management. Future research should aim to validate these results across more diverse populations and explore the integration of additional predictive factors that could refine the model's applicability. This study contributes to the computational health field by demonstrating how advanced machine learning techniques can be effectively applied in predicting health outcomes.

Hybrid modeling techniques for predicting chemical oxygen demand in wastewater treatment: a stacking ensemble learning approach with neural networks.

To ensure operational efficiency, promote sustainable wastewater treatment practices, and maintain compliance with environmental regulations, it is crucial to evaluate the parameters of treated effluent in wastewater treatment plants (WWTPs). Artificial neural network (ANN) analysis is a promising tool to predict the wastewater characteristics, as a substitute to tedious laboratory techniques. It enables proactive decision-making and contributes to the overall effectiveness of the treatment processes. The primary aim of this work is to develop a robust model for predicting chemical oxygen demand (COD), a key parameter for evaluating the operational efficiency of WWTPs. The research employed a dataset consisting of 527 samples and 22 features, derived from daily sensor readings in an urban WWTP. Additionally, to enhance the efficiency of the ANN framework for regression modeling, the dataset was augmented to 1054 samples using a robust synthetic data generator known as generative adversarial networks (GAN). In this study, K-means clustering combined with principal component analysis (PCA) is employed for feature selection and anomaly detection, aiming to enhance the regression model's performance by incorporating advanced ANN extensions, including polynomial, additive, and radial basis networks. Moreover, the model optimizes ANNs using advanced heuristic techniques such as genetic algorithms (GA), ant colony optimization (ACO), and particle swarm optimization (PSO). Furthermore, the performance of the models was assessed using the R2 (coefficient of determination), MSE (mean squared error), and loss value metrics along with visual performance indicators. The proposed model stacked ensemble method resulted in an MSE of 0.0012 and an R2 score of 0.95 and the GA_ANN model, despite undergoing optimization, achieved only an R2 score of 0.70, indicating considerable potential for improvement. In contrast, the ACO_ANN and PSO_ANN models showed significant performance boosts, with near-perfect R2 scores of about 0.98 and 0.99, respectively, making them the top-performing models overall. The results outlined in this article introduce an ANN stacking ensemble regression mode framework that can aid researchers in improving the operational efficiency of treatment plants.

18F-FDG PET/CT-based habitat radiomics combining stacking ensemble learning for predicting prognosis in hepatocellular carcinoma: a multi-center study

BackgroundThis study aims to develop habitat radiomic models to predict overall survival (OS) for hepatocellular carcinoma (HCC), based on the characterization of the intratumoral heterogeneity reflected in 18F-FDG PET/CT images.MethodsA total of 137 HCC patients from two institutions were retrospectively included. First, intratumoral habitats were achieved by a two-step unsupervised clustering process based on k-means clustering. Second, a total of 4032 radiomic features were extracted based on each habitat, including 2016 PET-based and 2016 CT-based radiomic features. Then, after feature selection, the stacking ensemble learning approach which combined six machine learning classifiers as the first-level learners with Cox proportional hazards regression as the second-level learner, was employed to build multiple radiomic models. Finally, the optimal model was selected based on the calculation of the C-index, and a combined model integrating with a clinical model was also constructed to identify the potentially complementary effect.ResultsThree spatially distinct habitats were identified in the two cohorts. Among a total of 30 stacking ensemble learning models established based on different combinations of 5 types of segmented volumes of interest (VOIs) with 6 types of classifiers, the MLP-Cox-habitat-2 model was selected as the optimal radiomic model with a C-index of 0.702 in the external validation cohort. Furthermore, the combined model integrating the optimal radiomic model with the clinical model achieved an improved C-index of 0.747. Consistently, the combined model outperformed the other models for OS prediction, with a time-dependent AUC of 0.835, 0.828, and 0.800 in the 1-year, 2-year, and 3-year OS, respectively.Conclusion18F-FDG PET/CT-based habitat radiomics outperformed traditional radiomics in OS prediction for HCC, with a further improved predictive power by integrating with the clinical model. The optimal combined habitat model was potentially promising in guiding individualized treatment for HCC.Trial registrationThis study was a retrospective study, so it was free from registration.

Forest plant invasions in the eastern United States: evidence of invasion debt in the wildland-urban interface

ContextNon-native invasive plants are a growing threat to forests. Meanwhile, the amount of forest within the wildland-urban interface (WUI) is increasing, with housing-associated disturbances enabling the spread of non-native plants.ObjectivesWe tested whether (1) WUI or non-WUI forests are more invaded, (2) WUI intermix forests (houses mingling with forest) are more invaded than interface forests (housing abutting forest), (3) invasion in WUI forests is delayed (invasion debt) following housing development, and (4) WUI forest invasion is associated with land cover context (a proxy for disturbance and propagule pressure).MethodsWe conducted statistical comparisons of plant invasion metrics using WUI status information intersected with ~ 45,000 forest inventory plots (collected ca. 2015). We evaluated potential drivers of invasion using an ensemble learning approach and adopted a mixed-effects modeling framework to assess relationships between drivers and invasion.ResultsOur analyses revealed that the degree of invasion in WUI forest plots was significantly higher. We found evidence for invasion debt, including greater invasion of plots in the WUI longer. WUI interface forests were more invaded than intermix forests. Agricultural and developed land cover at medium to large scales (~ 66 − 5,300 hectares) were most highly associated with WUI forest invasion.ConclusionsThese findings elucidate dynamic plant invasion patterns and processes occurring within the WUI. They underscore the importance of monitoring and managing forests that have recently entered the WUI, with an emphasis on sources of exotic plant propagules at relatively large scales, before the housing invasion debt comes due.

Hyperelastic and Stacked Ensemble-Driven Predictive Modeling of PEMFC Gaskets Under Thermal and Chemical Aging.

This study comprehensively investigates the stress distribution and aging effects in Ethylene Propylene Diene Monomer (EPDM) and Liquid Silicone Rubber (LSR) gasket materials through a novel integration of hyperelastic modeling and advanced machine learning techniques. By employing the Mooney-Rivlin, Ogden, and Yeoh hyperelastic models, we evaluated the mechanical behavior of EPDM and LSR under conditions of no aging, heat aging, and combined heat- and sulfuric-acid exposure. Each model revealed distinct sensitivities to stress distribution and material deformation, with peak von Mises stress values indicating that LSR experiences higher internal stress than EPDM across all conditions. For instance, without aging, LSR shows a von Mises stress of 24.17 MPa compared to 14.96 MPa for EPDM, while under heat and sulfuric acid exposure, LSR still exhibits higher stress values, showcasing its resilience under extreme conditions. Additionally, the ensemble learning approach achieved a classification accuracy of 98% for LSR and 84% for EPDM in predicting aging effects, underscoring the robustness of our predictive framework. These findings offer practical implications for selecting suitable gasket materials and developing predictive maintenance strategies in industrial applications, such as fuel cells, where material integrity under stress and aging is paramount.

Predicting future climate scenarios: a machine learning perspective on greenhouse gas emissions in agrifood systems

Global climate change is an extensive phenomenon characterized by alterations in weather patterns, temperature trends, and precipitation levels. These variations substantially impact agrifood systems, encompassing the interconnected components of farming, food production, and distribution. This article analyzes 8,100 data points with 27 input features that quantify diverse aspects of the agrifood system’s contribution to predicted Greenhouse Gas Emissions (GHGE). The study uses two machine learning algorithms, Long-Short Term Memory (LSTM) and Random Forest (RF), as well as a hybrid approach (LSTM-RF). The LSTM-RF model integrates the strengths of LSTM and RF. LSTMs are adept at capturing long-term dependencies in sequential data through memory cells, addressing the vanishing gradient problem. Meanwhile, with its ensemble learning approach, RF improves overall model performance and generalization by combining multiple weak learners. Additionally, RF provides insights into the importance of features, helping to understand the significant contributors to the model’s predictions. The results demonstrate that the LSTM-RF algorithm outperforms other algorithms (for the test subset, RMSE = 2.977 and R2 = 0.9990). These findings highlight the superior accuracy of the LSTM-RF algorithm compared to the individual LSTM and RF algorithms, with the RF algorithm being less accurate in comparison. As determined by Pearson correlation analysis, key variables such as on-farm energy use, pesticide manufacturing, and land use factors significantly influence GHGE outputs. Furthermore, this study uses a heat map to visually represent the correlation coefficient between the input variables and GHGE, enhancing our understanding of the complex interactions within the agrifood system. Understanding the intricate connection between climate change and agrifood systems is crucial for developing practices addressing food security and environmental challenges.

Ensemble Transfer Learning for Image Classification

The deep learning (DL) techniques used for image classification might not deliver the desired level of classification accuracy as some features belonging to some class of a dataset are missed during feature extraction. The ensemble learning (EL) based model improves classification accuracy by combining the strengths of individual classifiers. As a result, those features that were missed during feature extraction by a specific DL technique will be taken care of by another DL technique in an ensemble DL approach. In this paper, averaging EL (AENet), weighted averaging EL (WAENet), and stacking EL (StackedNet) approaches are proposed, considering the DenseNet201, EcientNetB0, and ResNetRS101 as base models. The predictions of the base models are averaged to generate the AENet. The WAENet is constructed by assigning weights to each base model based on their prediction and then taking their average. Similarly, the Stacked-Net is developed by considering the DenseNet201, EcientNetB0, and ResNetRS101 as base-learners and ResNetRS101 as meta-learner. Analysed performance of the considered pre-trained base models and the developed EL models on the standard and application-specific datasets such as MiniImageNet, CIFAR10, CIFAR100, Plant Village (PV), Tomato, Covid-19 and 9IndianFood. 80% of the datasets were used to train and 20% to test the base and proposed models. The models are trained for an epoch of 30, considering a learning rate of 0.001 and adam optimizer. The stackedNet delivered better results than others.

Conv‐ELSTM: An ensemble deep learning approach for predicting short‐term wind power

AbstractAccurate and reliable forecasting of wind power is essential for the stable integration of wind energy into the electrical grid. However, the chaotic nature of wind power presents a significant challenge in utilizing data for effective short‐term forecasting, such as 60‐min predictions. This article introduces a hybrid data‐driven framework that employs an ensemble deep learning model to provide highly precise short‐term wind power predictions. The framework leverages a data‐driven approach to identify the intrinsic components of wind power data, including high‐frequency and low‐frequency components. A convolutional layer‐based feature fusion network is then established to properly extract important information from irrelevant wind energy features. Subsequently, an ensemble of long short‐term memory (LSTM) networks is developed to forecast wind power using the fused features, thereby mitigating the disadvantage of a single prediction model. The numerical experiment is carried out based on two different real‐life datasets. The results demonstrate the effectiveness of the proposed method in forecasting short‐term wind power compared to five benchmarks.

Prospective and External Validation of an Ensemble Learning Approach to Sensitively Detect Intravenous Fluid Contamination in Basic Metabolic Panels.

Intravenous (IV) fluid contamination within clinical specimens causes an operational burden on the laboratory when detected, and potential patient harm when undetected. Even mild contamination is often sufficient to meaningfully alter results across multiple analytes. A recently reported unsupervised learning approach was more sensitive than routine workflows, but still lacked sensitivity to mild but significant contamination. Here, we leverage ensemble learning to more sensitively detect contaminated results using an approach which is explainable and generalizable across institutions. An ensemble-based machine learning pipeline of general and fluid-specific models was trained on real-world and simulated contamination and internally and externally validated. Benchmarks for performance assessment were derived from in silico simulations, in vitro experiments, and expert review. Fluid-specific regression models estimated contamination severity. SHapley Additive exPlanation (SHAP) values were calculated to explain specimen-level predictions, and algorithmic fairness was evaluated by comparing flag rates across demographic and clinical subgroups. The sensitivities, specificities, and Matthews correlation coefficients were 0.858, 0.993, and 0.747 for the internal validation set, and 1.00, 0.980, and 0.387 for the external set. SHAP values provided plausible explanations for dextrose- and ketoacidosis-related hyperglycemia. Flag rates from the pipeline were higher than the current workflow, with improved detection of contamination events expected to exceed allowable limits for measurement error and reference change values. An accurate, generalizable, and explainable ensemble-based machine learning pipeline was developed and validated for sensitively detecting IV fluid contamination. Implementing this pipeline would help identify errors that are poorly detected by current clinical workflows and a previously described unsupervised machine learning-based method.

Effective automatic classification methods via deep learning for myopic maculopathy.

Pathologic myopia (PM) associated with myopic maculopathy (MM) is a significant cause of visual impairment, especially in East Asia, where its prevalence has surged. Early detection and accurate classification of myopia-related fundus lesions are critical for managing PM. Traditional clinical analysis of fundus images is time-consuming and dependent on specialist expertise, driving the need for automated, accurate diagnostic tools. This study developed a deep learning-based system for classifying five types of MM using color fundus photographs. Five architectures-ResNet50, EfficientNet-B0, Vision Transformer (ViT), Contrastive Language-Image Pre-Training (CLIP), and RETFound-were utilized. An ensemble learning approach with weighted voting was employed to enhance model performance. The models were trained on a dataset of 2,159 annotated images from Shenzhen Eye Hospital, with performance evaluated using accuracy, sensitivity, specificity, F1-Score, Cohen's Kappa, and area under the receiver operating characteristic curve (AUC). The ensemble model achieved superior performance across all metrics, with an accuracy of 95.4% (95% CI: 93.0-97.0%), sensitivity of 95.4% (95% CI: 86.8-97.5%), specificity of 98.9% (95% CI: 97.1-99.5%), F1-Score of 95.3% (95% CI: 93.2-97.2%), Kappa value of 0.976 (95% CI: 0.957-0.989), and AUC of 0.995 (95% CI: 0.992-0.998). The voting ensemble method demonstrated robustness and high generalization ability in classifying complex lesions, outperforming individual models. The ensemble deep learning system significantly enhances the accuracy and reliability of MM classification. This system holds potential for assisting ophthalmologists in early detection and precise diagnosis, thereby improving patient outcomes. Future work could focus on expanding the dataset, incorporating image quality assessment, and optimizing the ensemble algorithm for better efficiency and broader applicability.

Stacked ensemble machine learning approach for electroencephalography based major depressive disorder classification using temporal statistics

Major depressive disorder (MDD) is a serious and widespread mental health condition that remains challenging to diagnose accurately. Traditional psychological assessments, which can be subjective and sometimes unreliable, emphasize the need for more objective diagnostic tools. In this study, we present a machine learning (ML) model designed to diagnose depression by analysing statistical time-domain features extracted from Electroencephalography (EEG) data. The model is built using a stacked ensemble ML approach, incorporating nine-base estimators with various meta-classifiers. Through multiple trials, the model achieved an accuracy of 98.01%, with precision and recall rates of 97.78% and 96.61%, respectively with Adaptive Boosting (AdaBoost) as the meta-classifer. We also investigated the effects of data sampling and the number of base classifiers on the model’s performance. The findings demonstrate that the stacked ensemble approach significantly enhances the accuracy of diagnosing MDD and that the proposed model outperforms the methods used in previous studies. This model offers a promising tool for psychologists and medical professionals to diagnose depression more reliably, potentially leading to better treatment outcomes for those affected by the disorder.

Experts' Cognition-Driven Ensemble Deep Learning for External Validation of Predicting Pathological Complete Response to Neoadjuvant Chemotherapy from Histological Images in Breast Cancer

In breast cancer, neoadjuvant chemotherapy (NAC) provides a standard treatment option for patients who have locally advanced cancer and some large operable tumors. A patient will have better prognosis when he has achieved a pathological complete response (pCR) with the treatment of NAC. There has been a trend to directly predict pCR to NAC from histological images based on deep learning (DL). However, the DL-based predictive models numerically have better performances in internal validation than in external validation. In this paper, we aim to alleviate this situation with an intrinsic approach. We propose an experts’ cognition-driven ensemble deep learning (ECDEDL) approach. Taking the cognition of both pathology and artificial intelligence experts into consideration to improve the generalization of the predictive model to the external validation, ECDEDL can intrinsically approximate the working paradigm of a human being which will refer to his various working experiences to make decisions. ECDEDL was validated with 695 WSIs collected from the same center as the primary dataset to develop the predictive model and perform the internal validation, and was also validated with 340 WSIs collected from other three centers as the external dataset to perform the external validation. In external validation, ECDEDL improves the AUCs of pCR prediction from 61.52(59.80-63.26) to 67.75(66.74-68.80) and the Accuracies of pCR prediction from 56.09(49.39-62.79) to 71.01(69.44-72.58). ECDEDL was quite effective for external validation of predicting pCR to NAC from histological images in breast cancer, numerically approximating the internal validation.

An ensemble deep learning model for classification of students as weak and strong learners via multiparametric analysis

Academic data predictions are significantly important for improving the overall education system's effectiveness by providing early identification of weak students and personalized learning strategies. This paper proposes a deep learning model for the identification of weak and strong students using ensemble learning and multiparametric analysis. It combines several machine learning algorithms, including Naive Bayes, Support Vector Machines, Multi-Layer Perceptron, and Logistic Regression using an ensemble learning approach to enhance the model’s performance. Additionally, a custom 1D Convolutional Neural Network (CNN) is designed for classification. It utilizes multiparametric analysis to identify weak and strong students considering various parameters such as age, academic performance, location, and online learning behavior. The evaluation results indicate the performance of the proposed model has been improved in comparison to MLA FIS, SHMM, and DRL by 16.5%, 5.5%, and 2.4%, in terms of precision, 16.4%, 6.5%, and 3.5 % in terms of accuracy and 10.4%, 2.5% and 6.5% in terms of recall. These improvisations described that the model is efficient for multidomain feature extraction, ensemble classification, and high-variance feature selection, which result in a deeper understanding of student performance.

Weighted ensemble deep learning approach for classification of gastrointestinal diseases in colonoscopy images aided by explainable AI

Gastrointestinal diseases are significant health issues worldwide, requiring early diagnosis due to their serious health implications. Therefore, detecting these diseases using artificial intelligence-based medical decision support systems through colonoscopy images plays a critical role in early diagnosis. In this study, a deep learning-based method is proposed for the classification of gastrointestinal diseases and colon anatomical landmarks using colonoscopy images. For this purpose, five different Convolutional Neural Network (CNN) models, namely Xception, ResNet-101, NASNet-Large, EfficientNet, and NASNet-Mobile, were trained. An ensemble model was created using class-based recall values derived from the validation performances of the top three models (Xception, ResNet-101, NASNet-Large). A user-friendly Graphical User Interface (GUI) was developed, allowing users to perform classification tasks and use Gradient-weighted Class Activation Mapping (Grad-CAM), an explainable AI tool, to visualize the regions from which the model derives information. Grad-CAM visualizations contribute to a better understanding of the model’s decision-making processes and play an important role in the application of explainable AI. In the study, eight labels, including anatomical markers such as z-line, pylorus, and cecum, as well as pathological findings like esophagitis, polyps, and ulcerative colitis, were classified using the KVASIR V2 dataset. The proposed ensemble model achieved a 94.125% accuracy on the KVASIR V2 dataset, demonstrating competitive performance compared to similar studies in the literature. Additionally, the precision and F1 score values ​​of this model are equal to 94.168% and 94.125%, respectively. These results suggest that the proposed method provides an effective solution for the diagnosis of GI diseases and can be beneficial for medical education.