Table 3 The research specifics on disease diagnosis based on medical imaging using ensemble AI
From: A systematic literature review: exploring the challenges of ensemble model for medical imaging
Authors (Year) | Algorithms | Modality | Performances | Domain | |
|---|---|---|---|---|---|
Baseline | Fusion | ||||
Ko et al. (2019) [23] | RetinaNet & Mask R-CNN | Weighted voting ensemble | 25,684 chest x-ray (CXR) images & 28,989 lung opacity bounding boxes | Mean average precision (mAP): 0.21746 | Pneumonia detection |
Livieris et al. (2019)[49] | SMO, C4.5, & kNN | Weighted voting ensemble | 5830 chest x-ray (Pneumonia), 566 lung mask (Tuberculosis), & 100 CT medical | Accuracy: Pneumonia 82.53–83.49%, Tuberculosis 69.79–71.73%, and CT medical 69–77% | Detecting lung abnormalities from chest X-rays |
Brunese et al. (2020)[46] | First order, shape, gray level co-occurrence matrix, gray level run length matrix & gray level size zone matrix | Weighted voting ensemble | 111,205 radiomic features extracted from MRI images | Accuracy of 99% | The benign grade I and the II, III, and IV malignant brain cancer detection |
Deb and Jha (2020)[24] | NASNet, MobileNet & DenseNet | Ensemble | Chest x-rays | Accuracy of 91.99% | COVID-19 detection into 3 classes: Community Acquired Pneumonia (CAP), normal, & COVID-19. |
Mao et al. (2020)[25] | RetinaNet & Mask R-CNN | Weighted average ensemble | 26,684 chest x-rays | Precision of 0.808, recall of 0.813, and mAP of 0.2283 | Detect pneumonia |
Suhail et al. (2020)[26] | Logistic regression & neural networks | Random forest ensemble | 287 cephalometric x-rays | Accuracy: primary or alternative outcome agreement 93–98% | Predicting orthodontic extractions |
Chandra et al.(2021) [27] | ANN, KNN, NB, DT, SVM-linear kernel, SVM-radial basis function, & SVM-polynomial kernel | Majority voting ensemble | 2346 chest x-rays | Accuracy of 93.41% | Automatic COVID screening (ACoS) system |
Golla et al.(2021) [44] | 2D and 3D versions of U-Net, V-Net, & DeepVesselNet | Ensemble | Abdominal CT scans | Dice similarity coefficients (DSC) of 0.758 for veins and 0.838 for arteries | Artery and vein segmentation |
Jiang et al. (2021) [45] | RRC model & 3D-CNN models | XGBoost ensemble | CT images: 7302 radiomic features and 17 radiological from 405 patients | AUROC 0.887–0.906 | Predicting MVI in HCC patients |
Jin et al. (2021) [28] | AlexNet, ReliefF, & SVM | Hybrid ensemble | 1743 chest x-rays | Overall accuracy rate: 98.642 ± 0.398% | Diagnosis between COVID-19, common viral pneumonia, and normal |
Rahman et al. (2021) [29] | ResNet101, VGG19, & DenseNet201 | XGBoost ensemble | 7000 chest x-rays | Accuracy of 99.92%, precision of 99.85%, & sensitivity of 100% | Detect tuberculosis (TB) and normal |
Sagor et al. (2021) [30] | Inception-v3, VGG16, & ResNet-50 | XGBoost ensemble | 112,120 chest x-rays | Accuracy of 88.14% | Detecting lung disease |
Tao et al. (2021) [14] | AlexNet, GoogleNet, & ResNet. | Relative majority voting ensemble | 7,500 lung CT images | Accuracy of 99.05% | Detecting COVID-19 between lung tumor and normal lung |
Bui et al. (2022) [31] | VGG, Resnet, Xception, & SVM | Majority voting ensemble | 95 Panoramic x-rays generate 533 tooth regions | Accuracy of 93.58%, sensitivity of 93.91%, & specificity of 93.33% | Automated caries screening |
Gokul et al. (2022) [32] | CAPSNet & VDSNet | Snapshot ensemble | 5,856 chest x-rays | Accuracy, precision, specificity: 0.98 & recall: 0.96 | Pneumonia diagnosis between pneumonia and normal class |
Imak et al. (2022) [33] | Multi-input AlexNet | Scored-based fusion ensemble | 340 periapical x-rays | Accuracy of 99.13% | Caries detection between caries and non caries class |
Iqball and Wani (2022) [34] | ResNet101, InceptionV3, MobileNetV2, NasNet, & Xception | Weighted sum ensemble | 1,088 chest x-rays | Accuracy, precision, recall: 100% | Pneumonia detection between COVID-19, pneumonia, normal |
Jaiswal et al. (2022) [35] | ResNet50V2, ResNet101V2, MobileNetV3Small, MobileNet, MobileNetV3Large, EfficientNetB0, EfficientNetB1, & EfficientNetB2 | XGBoost ensemble | 500 panoramic x-rays | Accuracy: tooth wear 91.8%, periapical 92.2%, periodontitis 92.4%, tooth decay 93.2%, missing tooth 91.6%, and impacted tooth 90.8% | Multi oral disease classification: tooth wear, periapical, periodontitis, tooth decay, missing tooth, and impacted tooth |
Muller et al. (2022) [50] | DenseNet121, EfficientNetB4, InceptionResNetV2, MobileNetV2, ResNeXt101, ResNet101, VGG16, Xception, & Vanilla architecture | Augmenting, bagging, & stacking ensemble | 5,000 CHMNIST (histology), 2,905 COVID (X-ray), 25,331 ISIC (dermoscopy), & 35,126 DRD (ophthalmoscopy) | Average accuracy: CHMNIST 98%, COVID 96%, ISIC 92.75%, & DRD 81.5% | Medical image classification: CHMNIST, COVID, ISIC, and DRD |
Alsubai (2023) [36] | Principle Component Analysis (PCA) & Chi-square (chi2) | Stacking ensemble (Extreme gradient boosting (XGB), random forest (RF), & extra trees classifier (ETC)) | 10,375 bitewing x-rays | Accuracy: 97.36%, precision: 96.14%, recall: 96,84%, & F1 score: 96.65% | Enhancing prediction of tooth caries and tooth normal |
Azhari et al. (2023) [37] | ResNet50, ResNext101, & Vgg19 | Ensemble U-Nets | 771 bitewing x-rays: adult 554 & pediatric 217 | IoU averaged of adult: zero 98%, primary 23%, moderate 19%, & advanced carious lesions 51% IoU averaged of pediatric: zero 97%, primary 8%, moderate 17%, & advanced carious lesions 25% | Detection of interproximal carious lesions on primary and permanent dentition |
Bhatt and Shah (2023) [38] | 3 CNN models with kernel sizes: 3×3, 5×5, & 7×7 | Weighted ensemble | 5,863 chest x-rays | Accuracy, recall, precision, & F1 score: 99.23% | Pneumonia detection between pneumonia and normal |
Haghanifar et al. (2023) [39] | CNN, InceptionNet, Encoder, & CheXNet, | PaXNet (Capsule network ensemble) | 470 panoramic x-rays where the teeth total is 1229; 616 maxillary & 613 mandibular | Accuracy: 81.44% maxilla & 73.67% mandible | Tooth segmentation and dental caries detection: healthy and carious |
Mabrouk et al. (2023) [40] | DenseNet169, MobileNetV2, & Vision Transformer | Ensemble learning | 5,856 chest x-rays | Accuracy: 93.91%, precision: 93.96%, recall: 92.99%, & F1 score: 93.43% | Pneumonia detection between pneumonia and normal |
Nakata and Siina (2023) [47] | 16 CNNs (Xception, InceptionV3, InceptionResNetV2, ResNet50 & 101, ResNeXt50 & 101, SeResNetXt50 & 101, EfficientnetB0-B6) | Soft voting, weighted average voting, weighted hard voting, & stacking ensemble | Ultrasound images: 6320 benign liver tumor (BLT), 2320 liver cyst (LCY), 9720 metastatic liver cancer (MLC), 7840 primary liver cancer (PLC) | Best ROC AUC: 0.944 BLT, 0.999 LCY, 0.891 MLC, & 0.903 PLC | Multiclass classification of ultrasound images of liver masses |
Paul and Naskar (2023) [41] | DenseNet169, VGG16, & InceptionV3 | Soft voting ensemble | 5,856 children’s lung chest x-ray | Accuracy of 92.79% | Children’s pneumonia detection between pneumonia and normal |
Tareq et al. (2023) [48] | YOLO v5s, v5m, v5l, & v5x, ResNet50, VGG16, & DenseNet3 | Ensemble | 233 de-identified anterior teeth specimens and 1703 after augmentation | Accuracy of 86.96% | Visual diagnostics of dental caries |
Dang et al. (2024) [51] | Deep neural networks (VGG16, ResNet34, & ResNet101) | Weighted ensemble | Medical image: 300 CVC-ColonDB, 912 CVC-EndoSceneStill-2017, 808 MICCAI2015, 1000 CAMUS-ED, & 1000 CAMUS-ES | Dice score: CVC-ColonDB 0.95, CVC-EndoSceneStill-2017 0.73, MICCAI2015 0.84, CAMUS-ED 0.94, & CAMUS-ES 0.94 | Two-layer ensemble of deep learning models for medical image segmentation |
Gupta et al. (2024) [42] | DenseNet201, MobileNetV2, & InceptionResNetV2 | Stacked ensemble | 5,216 chest x-ray | Accuracy of 94% | Precise pediatric pneumonia diagnosis between pneumonia and normal |
Marginean et al. (2024) [43] | U-Net, Feature Pyramid Network, & DeeplabV3 | Ensemble learning | 1,000 panoramic x-rays | Accuracy: 99.42% & mean dice coefficient 68.2% | Teeth segmentation and carious lesions segmentation |