<div><h3>Background</h3><div>Brain tumor classification and localization are essential for accurate diagnosis and effective treatment planning. With the increasing use of MRI in clinical workflows, automated tools have become crucial for assisting radiologists in providing fast and reliable analysis. Traditional approaches for tumor classification and segmentation are often time-consuming and subjective, underscoring the need for deep learning-based frameworks that can enhance diagnostic accuracy and efficiency.</div></div><div><h3>Method</h3><div>This study used the TCGA-GBM dataset from The Cancer Imaging Archive (TCIA) to assess a dual-stage deep learning framework combining a classification model with a RESUNET for segmentation. TCGA-GBM, part of the broader TCGA project, focuses on Glioblastoma Multiforme (GBM), a highly aggressive brain cancer. The dataset includes 3929 images, with 2556 non-tumor (class 0) and 1373 tumor (class 1) samples. The framework incorporated three convolutional neural network (CNN) architectures—MobileNet, NASNetMobile, and ResNet101—each enhanced with a Transfer Learning Layer for classification, followed by a RESUNET network for tumor localization. Transfer learning enabled the use of pre-trained weights, improving convergence speed and generalization. MobileNet offered a lightweight, efficient solution; NASNetMobile provided a strong balance between accuracy and computational cost; and ResNet101 delivered deeper feature extraction for higher precision. The RESUNET architecture, combining U-Net and residual learning, accurately segmented tumor regions, enabling effective integration of classification and localization within a unified framework.</div></div><div><h3>Results</h3><div>In the classification stage, the models achieved average accuracies of 0.9600 for MobileNet + Transfer Learning Layer, 0.9700 for NASNetMobile + Transfer Learning Layer, and 0.9500 for ResNet101 + Transfer Learning Layer, demonstrating their effectiveness in categorizing brain tumors. Performance was further evaluated using precision, recall, and F1 scores for both classes. MobileNet + Transfer Learning Layer achieved 0.95 precision, 0.99 recall, and 0.97 F1 for class 0, and 0.98, 0.92, and 0.95 for class 1. NASNetMobile + Transfer Learning Layer achieved 0.96, 0.99, and 0.98 for class 0, and 0.99, 0.93, and 0.96 for class 1. ResNet101 + Transfer Learning Layer achieved 0.97, 0.96, and 0.96 for class 0, and 0.93, 0.94, and 0.94 for class 1. For tumor localization, the RESUNET segmentation network accurately delineated tumor regions across all classification models.</div></div><div><h3>Conclusions</h3><div>The proposed dual-stage deep learning framework effectively automates both classification and localization of brain tumors from MRI scans. The results demonstrate strong performance across all architectures, with NASNetMobile + RESUNET achieving the most balanced combination of precision and efficiency. Compared to baseline CNNs and pr
{"title":"Dual-stage deep learning framework for brain tumor classification and localization using multimodal MRI scans","authors":"Deependra Rastogi , Prashant Johri , Sumit Singh Dhanda , Anand Singh , Suman Avdhesh Yadav , Arfat Ahmad Khan , Seifedine Kadry","doi":"10.1016/j.ibmed.2026.100361","DOIUrl":"10.1016/j.ibmed.2026.100361","url":null,"abstract":"<div><h3>Background</h3><div>Brain tumor classification and localization are essential for accurate diagnosis and effective treatment planning. With the increasing use of MRI in clinical workflows, automated tools have become crucial for assisting radiologists in providing fast and reliable analysis. Traditional approaches for tumor classification and segmentation are often time-consuming and subjective, underscoring the need for deep learning-based frameworks that can enhance diagnostic accuracy and efficiency.</div></div><div><h3>Method</h3><div>This study used the TCGA-GBM dataset from The Cancer Imaging Archive (TCIA) to assess a dual-stage deep learning framework combining a classification model with a RESUNET for segmentation. TCGA-GBM, part of the broader TCGA project, focuses on Glioblastoma Multiforme (GBM), a highly aggressive brain cancer. The dataset includes 3929 images, with 2556 non-tumor (class 0) and 1373 tumor (class 1) samples. The framework incorporated three convolutional neural network (CNN) architectures—MobileNet, NASNetMobile, and ResNet101—each enhanced with a Transfer Learning Layer for classification, followed by a RESUNET network for tumor localization. Transfer learning enabled the use of pre-trained weights, improving convergence speed and generalization. MobileNet offered a lightweight, efficient solution; NASNetMobile provided a strong balance between accuracy and computational cost; and ResNet101 delivered deeper feature extraction for higher precision. The RESUNET architecture, combining U-Net and residual learning, accurately segmented tumor regions, enabling effective integration of classification and localization within a unified framework.</div></div><div><h3>Results</h3><div>In the classification stage, the models achieved average accuracies of 0.9600 for MobileNet + Transfer Learning Layer, 0.9700 for NASNetMobile + Transfer Learning Layer, and 0.9500 for ResNet101 + Transfer Learning Layer, demonstrating their effectiveness in categorizing brain tumors. Performance was further evaluated using precision, recall, and F1 scores for both classes. MobileNet + Transfer Learning Layer achieved 0.95 precision, 0.99 recall, and 0.97 F1 for class 0, and 0.98, 0.92, and 0.95 for class 1. NASNetMobile + Transfer Learning Layer achieved 0.96, 0.99, and 0.98 for class 0, and 0.99, 0.93, and 0.96 for class 1. ResNet101 + Transfer Learning Layer achieved 0.97, 0.96, and 0.96 for class 0, and 0.93, 0.94, and 0.94 for class 1. For tumor localization, the RESUNET segmentation network accurately delineated tumor regions across all classification models.</div></div><div><h3>Conclusions</h3><div>The proposed dual-stage deep learning framework effectively automates both classification and localization of brain tumors from MRI scans. The results demonstrate strong performance across all architectures, with NASNetMobile + RESUNET achieving the most balanced combination of precision and efficiency. Compared to baseline CNNs and pr","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100361"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396425","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2025-12-22DOI: 10.1016/j.ibmed.2025.100336
Masoud Tabibian, Tahereh Razmpour, Rajib Saha
Effective lung cancer detection from CT scans remains critically challenged by class imbalance where benign and normal cases are underrepresented, leading to biased machine learning models with reduced sensitivity for minority classes and potentially missed diagnoses in cancer screening applications. We present a comprehensive comparative analysis of Diffusion Models and Deep Convolutional Generative Adversarial Networks (DCGANs), both incorporating modern architectural enhancements including spectral normalization, self-attention mechanisms, and conditional generation, for addressing class imbalance in lung cancer CT classification. Using the IQ-OTH/NCCD dataset comprising 1097 CT images across normal, benign, and malignant categories with statistical validation across 10 independent runs, we evaluated both approaches through quantitative image quality metrics (Fréchet Inception Distance, Kullback-Leibler divergence, Kernel Inception Distance, and Inception Score) and downstream classification performance. While Diffusion models consistently outperformed DCGANs across most image quality measures, the clinical significance was confirmed through task-based validation. Both generative approaches successfully addressed class imbalance: DCGAN-augmented datasets achieved overall accuracy of 0.9760 ± 0.0116 with benign recall improvement from 0.833 to 0.933, while Diffusion-augmented datasets reached superior performance of 0.9959 ± 0.0068 with perfect benign recall (1.000 ± 0.000). Critically for cancer screening where false negatives carry severe consequences, Diffusion maintained the highest malignant detection sensitivity (0.997 ± 0.008) with substantially lower performance variance, demonstrating more consistent synthetic data quality. These findings establish that while both modern architectures can mitigate class imbalance, Diffusion models' superior recall performance and lower variability position them as the preferred approach for high-stakes clinical applications, demonstrating that ultimate validation must prioritize downstream clinical task performance over image quality metrics alone.
{"title":"Diffusion models vs. DCGANs for class-imbalanced lung cancer CT classification: A comparative study","authors":"Masoud Tabibian, Tahereh Razmpour, Rajib Saha","doi":"10.1016/j.ibmed.2025.100336","DOIUrl":"10.1016/j.ibmed.2025.100336","url":null,"abstract":"<div><div>Effective lung cancer detection from CT scans remains critically challenged by class imbalance where benign and normal cases are underrepresented, leading to biased machine learning models with reduced sensitivity for minority classes and potentially missed diagnoses in cancer screening applications. We present a comprehensive comparative analysis of Diffusion Models and Deep Convolutional Generative Adversarial Networks (DCGANs), both incorporating modern architectural enhancements including spectral normalization, self-attention mechanisms, and conditional generation, for addressing class imbalance in lung cancer CT classification. Using the IQ-OTH/NCCD dataset comprising 1097 CT images across normal, benign, and malignant categories with statistical validation across 10 independent runs, we evaluated both approaches through quantitative image quality metrics (Fréchet Inception Distance, Kullback-Leibler divergence, Kernel Inception Distance, and Inception Score) and downstream classification performance. While Diffusion models consistently outperformed DCGANs across most image quality measures, the clinical significance was confirmed through task-based validation. Both generative approaches successfully addressed class imbalance: DCGAN-augmented datasets achieved overall accuracy of 0.9760 ± 0.0116 with benign recall improvement from 0.833 to 0.933, while Diffusion-augmented datasets reached superior performance of 0.9959 ± 0.0068 with perfect benign recall (1.000 ± 0.000). Critically for cancer screening where false negatives carry severe consequences, Diffusion maintained the highest malignant detection sensitivity (0.997 ± 0.008) with substantially lower performance variance, demonstrating more consistent synthetic data quality. These findings establish that while both modern architectures can mitigate class imbalance, Diffusion models' superior recall performance and lower variability position them as the preferred approach for high-stakes clinical applications, demonstrating that ultimate validation must prioritize downstream clinical task performance over image quality metrics alone.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100336"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145926916","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2026-01-12DOI: 10.1016/j.ibmed.2026.100346
Ahmed Kateb Jumaah Al-Nussairi , Saleem Malik , Yasser Taha Alzubaidi , S Gopal Krishna Patro , Kasim Sakran Abass , Iman Basheti , Mohammad Khishe
Most AI breast cancer detection systems use single-modality imaging algorithms, limiting clinical reliability. Early and accurate detection improves therapy and mortality. These challenges are addressed by Transformer-driven multi-modal fusion and explainable deep learning system, TransFusion-BCNet for breast cancer diagnosis. The framework consists of three parts. The TriFusion-Transformer (TriFT) performs three-tier fusion: intra-modality fusion across multiple mammogram views and imaging sources, inter-modality fusion combining mammogram, ultrasound, MRI, and clinical features, and decision-level fusion for robust outcome prediction. TriFT detects complicated connections across heterogeneous modalities, unlike classical fusion. Second, we present the FusionAttribution Map (FAMap), a dual-level interpretability mechanism that generates imaging data region-level saliency maps and modality-level contribution scores to evaluate input source influence. This openness helps clinicians understand where and which modality drives predictions. Third, the MetaFusion Optimizer (MFO) adjusts fusion weights, network depth, and learning parameters via evolutionary search and gradient-based fine-tuning. Traditional optimizers lack model generalization and training stability. This staged technique improved both. TransFusion-BCNet outperforms CNN–Transformer hybrids with 99.4 % accuracy, 99.0 % precision, 99.2 % recall, and 99.1 % F1-score in extensive CBIS-DDSM,BUSI, TCGA-BRCA and RIDER Breast MRI datasets. With TriFT, FAMap, and MFO, TransFusion-BCNet provides a robust, transparent, and clinically interpretable diagnostic framework, improving AI in breast cancer screening and decision assistance.
{"title":"TransFusion-BCNet: A transformer-driven multi-modal fusion and explainable deep learning framework for breast cancer diagnosis","authors":"Ahmed Kateb Jumaah Al-Nussairi , Saleem Malik , Yasser Taha Alzubaidi , S Gopal Krishna Patro , Kasim Sakran Abass , Iman Basheti , Mohammad Khishe","doi":"10.1016/j.ibmed.2026.100346","DOIUrl":"10.1016/j.ibmed.2026.100346","url":null,"abstract":"<div><div>Most AI breast cancer detection systems use single-modality imaging algorithms, limiting clinical reliability. Early and accurate detection improves therapy and mortality. These challenges are addressed by Transformer-driven multi-modal fusion and explainable deep learning system, TransFusion-BCNet for breast cancer diagnosis. The framework consists of three parts. The TriFusion-Transformer (TriFT) performs three-tier fusion: intra-modality fusion across multiple mammogram views and imaging sources, inter-modality fusion combining mammogram, ultrasound, MRI, and clinical features, and decision-level fusion for robust outcome prediction. TriFT detects complicated connections across heterogeneous modalities, unlike classical fusion. Second, we present the FusionAttribution Map (FAMap), a dual-level interpretability mechanism that generates imaging data region-level saliency maps and modality-level contribution scores to evaluate input source influence. This openness helps clinicians understand where and which modality drives predictions. Third, the MetaFusion Optimizer (MFO) adjusts fusion weights, network depth, and learning parameters via evolutionary search and gradient-based fine-tuning. Traditional optimizers lack model generalization and training stability. This staged technique improved both. TransFusion-BCNet outperforms CNN–Transformer hybrids with 99.4 % accuracy, 99.0 % precision, 99.2 % recall, and 99.1 % F1-score in extensive CBIS-DDSM,BUSI, TCGA-BRCA and RIDER Breast MRI datasets. With TriFT, FAMap, and MFO, TransFusion-BCNet provides a robust, transparent, and clinically interpretable diagnostic framework, improving AI in breast cancer screening and decision assistance.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100346"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146077478","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Reliable identification of nasal bone fractures on lateral radiographs is a persistent clinical challenge, largely due to the subtle radiographic presentation of fractures, variability and noise inherent in routine imaging, and the limited availability of expert-annotated datasets. To overcome these obstacles, we present MoE-Net, a precision-oriented Mixture-of-Experts Network designed to enhance automated nasal bone fracture detection. The proposed approach systematically evaluates a pool of ten pre-trained convolutional neural network (CNN) and transformer-based architectures, from which the three most accurate and error-resilient models, InceptionResNetV2, DenseNet121, and Xception, are selected as specialized experts. Each expert model undergoes targeted fine-tuning, while a genetic algorithm optimizes their contribution weights within the ensemble to maximize predictive performance. MoE-Net demonstrates substantial performance gains over individual models and existing handcrafted-feature-based methods. On the test dataset, the framework achieves an accuracy of 91.70%, sensitivity of 91.70%, specificity of 89.95%, F2-score of 91.69%, Matthews Correlation Coefficient of 83.48%, and an area under the ROC curve of 91.99%. This performance reflects a clinically meaningful balance between minimizing false negatives, critical for preventing missed diagnoses, and controlling false positives to avoid unnecessary interventions. These findings support the clinical applicability of MoE-Net as a robust, high-performance decision-support tool for nasal bone fracture detection. The study highlights the advantages of precision-driven ensemble strategies in medical imaging and underscores their potential to improve diagnostic accuracy and contribute to more efficient patient care.
{"title":"MoE-Net: A deep ensemble framework optimized by genetic algorithm for nasal bone fracture detection on lateral X-ray images","authors":"Mobin Mehrpour , Seyed Abolghasem Mirroshandel , Tahereh Mortezaei , Zahra Dalili Kajan","doi":"10.1016/j.ibmed.2026.100359","DOIUrl":"10.1016/j.ibmed.2026.100359","url":null,"abstract":"<div><div>Reliable identification of nasal bone fractures on lateral radiographs is a persistent clinical challenge, largely due to the subtle radiographic presentation of fractures, variability and noise inherent in routine imaging, and the limited availability of expert-annotated datasets. To overcome these obstacles, we present MoE-Net, a precision-oriented Mixture-of-Experts Network designed to enhance automated nasal bone fracture detection. The proposed approach systematically evaluates a pool of ten pre-trained convolutional neural network (CNN) and transformer-based architectures, from which the three most accurate and error-resilient models, InceptionResNetV2, DenseNet121, and Xception, are selected as specialized experts. Each expert model undergoes targeted fine-tuning, while a genetic algorithm optimizes their contribution weights within the ensemble to maximize predictive performance. MoE-Net demonstrates substantial performance gains over individual models and existing handcrafted-feature-based methods. On the test dataset, the framework achieves an accuracy of 91.70%, sensitivity of 91.70%, specificity of 89.95%, F2-score of 91.69%, Matthews Correlation Coefficient of 83.48%, and an area under the ROC curve of 91.99%. This performance reflects a clinically meaningful balance between minimizing false negatives, critical for preventing missed diagnoses, and controlling false positives to avoid unnecessary interventions. These findings support the clinical applicability of MoE-Net as a robust, high-performance decision-support tool for nasal bone fracture detection. The study highlights the advantages of precision-driven ensemble strategies in medical imaging and underscores their potential to improve diagnostic accuracy and contribute to more efficient patient care.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100359"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188803","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Predicting chronic disease remains a crucial problem, particularly in low-resource environments where accurate and timely predictions are of utmost importance. Machine learning methods do not adequately generalize, encounter data imbalance and computation issues when used. In this study, an improved model that makes predictions using Synthetic minority over-sampling technique and edited nearest neighbor balancing techniques and Principal Component Analysis and Gradient Boosting to predict chronic diseases based on readily available clinical profiles is proposed. The dataset used in this study was retrieved from provide precise source: Kaggle dataset. It contains over 1500 anonymized patient records and includes 15 features such as demographic, lifestyle, and clinical measures. Standardized encoding labels, and adjusting classes using SMOTE-ENN have been completed before the PCA was conducted to improve computation speed and reduce overfitting. Decision Tree, Random Forest, LightGBM, XGBoost have been used for comparison to suggest the best performing model seen to be Gradient Boosting. PCA is performed, and the Gradient Boosting approach produces better results. Precision measures how often the classification system is correct when making a positive test result, while recall is determined using a contingency table, and the F1 score, the possibility of modeling outcomes out of 100 trials. The model proposed in the experiment provides the following outputs: Accuracy (CV: 99.33 %, CI: 98.90 %–99.50 %), Precision (CV: 99 %, CI: 98 %–99.5 %), Recall (CV: 99 %, CI: 98 %–99.5 %), F1-Score (CV: 99 %, CI: 98 %–99.5 %). The model's performance was evaluated using cross-validation, yielding an accuracy of 98.90 %. The classifying system performance is specified by the ROC-AUC ranking. It outperforms the model making indefinite projections; its ROC-AUC value is greater than 0.99. The suggested model is a robust, interpretable, and high-precision approach for the early detection of chronic conditions. Therefore, the suggested machine learning system can deliver a considerable promise with respect to creating patient-oriented outcomes.
{"title":"Enhancing risk prediction for diabetes, hypertension, and heart disease using SMOTE-ENN balancing with PCA and gradient boosting in healthcare AI","authors":"Tapon Paul , Md Assaduzzaman , Nafiz Fahad , Md Jakir Hossen","doi":"10.1016/j.ibmed.2025.100339","DOIUrl":"10.1016/j.ibmed.2025.100339","url":null,"abstract":"<div><div>Predicting chronic disease remains a crucial problem, particularly in low-resource environments where accurate and timely predictions are of utmost importance. Machine learning methods do not adequately generalize, encounter data imbalance and computation issues when used. In this study, an improved model that makes predictions using Synthetic minority over-sampling technique and edited nearest neighbor balancing techniques and Principal Component Analysis and Gradient Boosting to predict chronic diseases based on readily available clinical profiles is proposed. The dataset used in this study was retrieved from provide precise source: Kaggle dataset. It contains over 1500 anonymized patient records and includes 15 features such as demographic, lifestyle, and clinical measures. Standardized encoding labels, and adjusting classes using SMOTE-ENN have been completed before the PCA was conducted to improve computation speed and reduce overfitting. Decision Tree, Random Forest, LightGBM, XGBoost have been used for comparison to suggest the best performing model seen to be Gradient Boosting. PCA is performed, and the Gradient Boosting approach produces better results. Precision measures how often the classification system is correct when making a positive test result, while recall is determined using a contingency table, and the F1 score, the possibility of modeling outcomes out of 100 trials. The model proposed in the experiment provides the following outputs: Accuracy (CV: 99.33 %, CI: 98.90 %–99.50 %), Precision (CV: 99 %, CI: 98 %–99.5 %), Recall (CV: 99 %, CI: 98 %–99.5 %), F1-Score (CV: 99 %, CI: 98 %–99.5 %). The model's performance was evaluated using cross-validation, yielding an accuracy of 98.90 %. The classifying system performance is specified by the ROC-AUC ranking. It outperforms the model making indefinite projections; its ROC-AUC value is greater than 0.99. The suggested model is a robust, interpretable, and high-precision approach for the early detection of chronic conditions. Therefore, the suggested machine learning system can deliver a considerable promise with respect to creating patient-oriented outcomes.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100339"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146037648","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2026-01-07DOI: 10.1016/j.ibmed.2026.100345
Roopitha C.H , Veena Mayya , V. Sivakumar , Vathsala Patil , Divya Pai , Pavithra Varchas
Timely identification of impacted canines is crucial for preventing complications such as root resorption, misalignment, and damage to neighboring teeth. This research evaluates image processing techniques for AI-based oral decision support systems using orthopantomogram (OPG) images to determine effective preprocessing methods to improve deep learning (DL) model performance. The approach involved preprocessing OPG images followed by object detection using YOLOv5. Preprocessing techniques included median filtering, Gaussian blur, CLAHE, image sharpening, histogram stretching, and CLAHE combined with image sharpening. Performance was evaluated using precision, recall, F1 score, mAP, and IoU metrics. The CLAHE-sharpening combination achieved superior performance with precision of 0.934, recall of 0.932, F1 score of 0.931, mAP of 0.948, and IoU of 0.741, significantly outperforming unprocessed images. Grad-CAM visualizations confirmed that preprocessing enabled the model to identify relevant regions effectively. This study emphasizes the importance of preprocessing methods in improving the diagnostic accuracy in dental radiography for improved treatment planning.
{"title":"Enhancing AI-based oral decision support systems: Hybrid image processing for detecting impacted maxillary canines in orthopantomograms","authors":"Roopitha C.H , Veena Mayya , V. Sivakumar , Vathsala Patil , Divya Pai , Pavithra Varchas","doi":"10.1016/j.ibmed.2026.100345","DOIUrl":"10.1016/j.ibmed.2026.100345","url":null,"abstract":"<div><div>Timely identification of impacted canines is crucial for preventing complications such as root resorption, misalignment, and damage to neighboring teeth. This research evaluates image processing techniques for AI-based oral decision support systems using orthopantomogram (OPG) images to determine effective preprocessing methods to improve deep learning (DL) model performance. The approach involved preprocessing OPG images followed by object detection using YOLOv5. Preprocessing techniques included median filtering, Gaussian blur, CLAHE, image sharpening, histogram stretching, and CLAHE combined with image sharpening. Performance was evaluated using precision, recall, F1 score, mAP, and IoU metrics. The CLAHE-sharpening combination achieved superior performance with precision of 0.934, recall of 0.932, F1 score of 0.931, mAP of 0.948, and IoU of 0.741, significantly outperforming unprocessed images. Grad-CAM visualizations confirmed that preprocessing enabled the model to identify relevant regions effectively. This study emphasizes the importance of preprocessing methods in improving the diagnostic accuracy in dental radiography for improved treatment planning.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100345"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145926930","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Breast cancer is a leading global cause of cancer-related mortality, where early diagnosis is essential for improved survival outcomes. Although deep learning has shown strong performance in pathology image classification, many models remain difficult to interpret, which limits clinical trust and practical adoption. This study aimed to develop an explainable deep learning framework for classifying breast tissue based on Biglycan (BGN) biomarker expression.
Methods
Immunohistochemical photomicrographs from the publicly available Biglycan Breast Cancer Dataset were processed and classified into cancerous and healthy categories. Three transfer learning models, EfficientNet-B0, DenseNet-161, and ResNet-50, were fine-tuned using ImageNet pre-trained weights under a unified training setting with the Adam optimizer (learning rate = 0.001, batch size = 32, epochs = 50). Model performance was evaluated using accuracy, precision, recall, and F1-score. To enhance interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize tissue regions that contributed most to the model's predictions.
Results
EfficientNet-B0 achieved the best overall performance with 99 % accuracy, precision of 0.97, recall of 1.00 for the cancerous class, and an F1-score of 0.99. Grad-CAM heatmaps indicated that the model focused on diagnostically relevant tissue regions associated with strong BGN-related staining patterns. The best-performing model was integrated into a lightweight web-based application to enable real-time image upload and prediction.
Conclusion
This study presents an explainable transfer learning approach for BGN-driven breast tissue classification with strong performance on the Biglycan dataset. The integration of Grad-CAM provides region-level visual explanations that improve transparency, while the web deployment demonstrates a practical pathway for accessible decision support in digital pathology.
{"title":"Explainable AI for breast cancer detection: Biglycan biomarker classification with transfer learning","authors":"Md. Mominul Islam , Naime Akter , Md. Assaduzzaman , Md. Monir Hossain Shimul , Rahmatul Kabir Rasel Sarker","doi":"10.1016/j.ibmed.2025.100340","DOIUrl":"10.1016/j.ibmed.2025.100340","url":null,"abstract":"<div><h3>Background</h3><div>Breast cancer is a leading global cause of cancer-related mortality, where early diagnosis is essential for improved survival outcomes. Although deep learning has shown strong performance in pathology image classification, many models remain difficult to interpret, which limits clinical trust and practical adoption. This study aimed to develop an explainable deep learning framework for classifying breast tissue based on Biglycan (BGN) biomarker expression.</div></div><div><h3>Methods</h3><div>Immunohistochemical photomicrographs from the publicly available Biglycan Breast Cancer Dataset were processed and classified into cancerous and healthy categories. Three transfer learning models, EfficientNet-B0, DenseNet-161, and ResNet-50, were fine-tuned using ImageNet pre-trained weights under a unified training setting with the Adam optimizer (learning rate = 0.001, batch size = 32, epochs = 50). Model performance was evaluated using accuracy, precision, recall, and F1-score. To enhance interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize tissue regions that contributed most to the model's predictions.</div></div><div><h3>Results</h3><div>EfficientNet-B0 achieved the best overall performance with 99 % accuracy, precision of 0.97, recall of 1.00 for the cancerous class, and an F1-score of 0.99. Grad-CAM heatmaps indicated that the model focused on diagnostically relevant tissue regions associated with strong BGN-related staining patterns. The best-performing model was integrated into a lightweight web-based application to enable real-time image upload and prediction.</div></div><div><h3>Conclusion</h3><div>This study presents an explainable transfer learning approach for BGN-driven breast tissue classification with strong performance on the Biglycan dataset. The integration of Grad-CAM provides region-level visual explanations that improve transparency, while the web deployment demonstrates a practical pathway for accessible decision support in digital pathology.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100340"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145926928","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Accurate evaluation of HER2 gene amplification is critical for guiding breast cancer treatment decisions. This study proposes a deep learning-based diagnostic system for analyzing Dual In Situ Hybridization (DISH) images to support HER2 status assessment. The system integrates two models— YOLOv11-seg for cell detection and YOLOv11 object detection models for HER2 and CEP17 signal quantification—into a unified pipeline. High-resolution whole-slide images were preprocessed and annotated to train the models, which were then embedded into a standalone application designed for clinical environments. Upon uploading TIFF format images, the application performs automated cell detection, red/black signal analysis, and HER2/CEP17 ratio computation. Experimental results demonstrated an accuracy 95.24 % for the best identification and mean deviations of 6.08 % (CEP17) and 12.78 % (HER2) compared with manual counting. Statistical analyses confirm high consistency, particularly in red signal detection. Clinical feedback under scores the system's ease of use, accuracy, and potential to reduce diagnostic burden. The proposed approach demonstrates strong feasibility for routine adoption in pathology workflows.
{"title":"Automated detection of HER2 gene copy number in breast cancer using deep learning techniques","authors":"Terisara Micaraseth , Shanop Shuangshoti , Sakdina Prommaouan , Somruetai Shuangshoti , Rizwan Ullah , Gridsada Phanomchoeng","doi":"10.1016/j.ibmed.2025.100333","DOIUrl":"10.1016/j.ibmed.2025.100333","url":null,"abstract":"<div><div>Accurate evaluation of HER2 gene amplification is critical for guiding breast cancer treatment decisions. This study proposes a deep learning-based diagnostic system for analyzing Dual In Situ Hybridization (DISH) images to support HER2 status assessment. The system integrates two models— YOLOv11-seg for cell detection and YOLOv11 object detection models for HER2 and CEP17 signal quantification—into a unified pipeline. High-resolution whole-slide images were preprocessed and annotated to train the models, which were then embedded into a standalone application designed for clinical environments. Upon uploading TIFF format images, the application performs automated cell detection, red/black signal analysis, and HER2/CEP17 ratio computation. Experimental results demonstrated an accuracy 95.24 % for the best identification and mean deviations of 6.08 % (CEP17) and 12.78 % (HER2) compared with manual counting. Statistical analyses confirm high consistency, particularly in red signal detection. Clinical feedback under scores the system's ease of use, accuracy, and potential to reduce diagnostic burden. The proposed approach demonstrates strong feasibility for routine adoption in pathology workflows.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100333"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145926924","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Eosinophilic granulomatosis with polyangiitis (EGPA), formerly Churg-Strauss syndrome, is a rare systemic vasculitis often diagnosed late due to its heterogeneous presentation, leading to severe complications—particularly cardiac involvement, a major cause of morbidity and mortality. We developed EGPA-ML, an artificial intelligence (AI)-based tool using supervised machine learning (ML), to support early and accurate EGPA diagnosis, especially in non-specialized settings.
Methods
A retrospective cohort of patients evaluated for suspected vasculitis at Hedi Chaker Hospital, Sfax, Tunisia, from 1997 to 2023 (nearly three decades), provided 1904 clinical, biological, and histological features. After data cleaning, standardization, and feature selection, 56 key features were retained. Patients were classified as {EGPA} or {NOT_EGPA} per the 2022 ACR/EULAR criteria, with expert consensus (κ = 0.85). Multiple supervised ML algorithms were evaluated via 10-fold cross-validation. The best model was integrated into EGPA-ML, a Java-based clinical decision support system. Performance was assessed on an independent dataset of n = 280 key features, with reference classification {EGPA}/{NOT_EGPA} validated by experts (κ = 0.89).
Results
On the test and evaluation dataset, EGPA-ML achieved a recall of 0.992, precision of 0.869, and F1-score of 0.926. Feature importance analysis identified asthma and eosinophil count as top predictors (36.5 % each), followed by ANCA status, vascular purpura, and histological vasculitis.
Conclusions
EGPA-ML is a high-performance, interpretable, and adaptive tool based on supervised ML, supporting timely EGPA diagnosis. It represents a practical advancement for clinical decision-making in rare diseases, particularly in internal medicine, pulmonology, and cardiology.
{"title":"A machine learning–based method for supporting the diagnosis of eosinophilic granulomatosis with polyangiitis: Development and evaluation","authors":"Yosra Bouattour , Mohamed Hédi Maâloul , Zouhir Bahloul , Sameh Marzouk","doi":"10.1016/j.ibmed.2025.100317","DOIUrl":"10.1016/j.ibmed.2025.100317","url":null,"abstract":"<div><h3>Background/objectives</h3><div>Eosinophilic granulomatosis with polyangiitis (EGPA), formerly Churg-Strauss syndrome, is a rare systemic vasculitis often diagnosed late due to its heterogeneous presentation, leading to severe complications—particularly cardiac involvement, a major cause of morbidity and mortality. We developed EGPA-ML, an artificial intelligence (AI)-based tool using supervised machine learning (ML), to support early and accurate EGPA diagnosis, especially in non-specialized settings.</div></div><div><h3>Methods</h3><div>A retrospective cohort of patients evaluated for suspected vasculitis at Hedi Chaker Hospital, Sfax, Tunisia, from 1997 to 2023 (nearly three decades), provided 1904 clinical, biological, and histological features. After data cleaning, standardization, and feature selection, 56 key features were retained. Patients were classified as {EGPA} or {NOT_EGPA} per the 2022 ACR/EULAR criteria, with expert consensus (κ = 0.85). Multiple supervised ML algorithms were evaluated via 10-fold cross-validation. The best model was integrated into EGPA-ML, a Java-based clinical decision support system. Performance was assessed on an independent dataset of n = 280 key features, with reference classification {EGPA}/{NOT_EGPA} validated by experts (κ = 0.89).</div></div><div><h3>Results</h3><div>On the test and evaluation dataset, EGPA-ML achieved a recall of 0.992, precision of 0.869, and F1-score of 0.926. Feature importance analysis identified asthma and eosinophil count as top predictors (36.5 % each), followed by ANCA status, vascular purpura, and histological vasculitis.</div></div><div><h3>Conclusions</h3><div>EGPA-ML is a high-performance, interpretable, and adaptive tool based on supervised ML, supporting timely EGPA diagnosis. It represents a practical advancement for clinical decision-making in rare diseases, particularly in internal medicine, pulmonology, and cardiology.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100317"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145926989","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2026-02-28DOI: 10.1016/j.ibmed.2026.100367
Salomon Massoda , Claudia Émond , Vincent Tellier , Hossein Kourkchi , Hind Rakkay , Alexandre Sasseville , George Stoica , Albert Chau , Stuart Coupland , Claude Hariton
Background
We integrated the features used to generate retinal signals, as well as the measured outcome, in our high-density signal concept. In this work, we expand to time-frequency signal analysis processes, and present a novel approach based upon specifically designed frequency patterns.
Methods
A multicentre clinical study was conducted to support prediction models able to accurately differentiate two serious mental health conditions: schizophrenia and type 1 bipolar disorder. Time-frequency domain features were processed using sparse representation (SR) and Principal Component Analysis (PCA). Commonly used mother wavelets have been chosen, as well as specific mother wavelets designed from reference retinal signal patterns. Wavelet coefficients and spectral entropies were selected as retinal signal features in developing the prediction models.
Results
Mappings of predictors in the time-frequency domain demonstrate that retinal signal regions that were not yet investigated include meaningful classifiers. With the implementation of SR combined with time-frequency retinal signal analysis, we developed classification models achieving meaningful level of performances. Cross-validation analyses with 100 replicates reached the highest training performance (99% mean accuracy) with SVM analysis using Gaus8 predictors. In testing, cross-validation highest performance was obtained with the Ridge logistic regression using the Gaus2 predictors (91% mean accuracy). In prediction models using ERGW or ERGWB wavelet coefficients as predictors, cross-validation reached highest testing performance (89% mean accuracy) with LASSO logistic regression model when using ERGWB real even wavelet coefficients as predictors.
Conclusion
Our approach leads to a more refined multimodal construct, able to differentiate subtle information within biosignatures in complex neuropsychiatric disorders. One distinctive improvement is the design and application of pattern-adapted wavelets derived from control-group retinal signals, to be used as deciphering tools for biosignature extraction in pathological conditions.
{"title":"Sparse representation of high-density retinal signal in time-frequency domain to support diagnosis in psychiatric disorders","authors":"Salomon Massoda , Claudia Émond , Vincent Tellier , Hossein Kourkchi , Hind Rakkay , Alexandre Sasseville , George Stoica , Albert Chau , Stuart Coupland , Claude Hariton","doi":"10.1016/j.ibmed.2026.100367","DOIUrl":"10.1016/j.ibmed.2026.100367","url":null,"abstract":"<div><h3>Background</h3><div>We integrated the features used to generate retinal signals, as well as the measured outcome, in our high-density signal concept. In this work, we expand to time-frequency signal analysis processes, and present a novel approach based upon specifically designed frequency patterns.</div></div><div><h3>Methods</h3><div>A multicentre clinical study was conducted to support prediction models able to accurately differentiate two serious mental health conditions: schizophrenia and type 1 bipolar disorder. Time-frequency domain features were processed using sparse representation (SR) and Principal Component Analysis (PCA). Commonly used mother wavelets have been chosen, as well as specific mother wavelets designed from reference retinal signal patterns. Wavelet coefficients and spectral entropies were selected as retinal signal features in developing the prediction models.</div></div><div><h3>Results</h3><div>Mappings of predictors in the time-frequency domain demonstrate that retinal signal regions that were not yet investigated include meaningful classifiers. With the implementation of SR combined with time-frequency retinal signal analysis, we developed classification models achieving meaningful level of performances. Cross-validation analyses with 100 replicates reached the highest training performance (99% mean accuracy) with SVM analysis using Gaus8 predictors. In testing, cross-validation highest performance was obtained with the Ridge logistic regression using the Gaus2 predictors (91% mean accuracy). In prediction models using ERGW or ERGWB wavelet coefficients as predictors, cross-validation reached highest testing performance (89% mean accuracy) with LASSO logistic regression model when using ERGWB real even wavelet coefficients as predictors.</div></div><div><h3>Conclusion</h3><div>Our approach leads to a more refined multimodal construct, able to differentiate subtle information within biosignatures in complex neuropsychiatric disorders. One distinctive improvement is the design and application of pattern-adapted wavelets derived from control-group retinal signals, to be used as deciphering tools for biosignature extraction in pathological conditions.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"13 ","pages":"Article 100367"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396420","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：