Table 6. Hemorrhages detection approaches.
| Year | Reference | Methods | Database | Results with evaluation metrics |
|---|---|---|---|---|
| 2024 | Sathiyaseelan et al. (86) |
Fast-CNN, modified U-Net, enhanced machine-based diagnostic | IDRiD, DIARETDB1 | Specificity: 99.6%; sensitivity: 80%; accuracy: 98.6% |
| 2024 | Biju and Shanthi (87) | Enhanced visual geometry group model, data augmentation, Gabor transform, hemorrhages segmentation, and classification module | HRF, DIARETDB1 | Specificity: 98.71%, 98.44%; sensitivity: 98.59%, 98.37%; accuracy: 98.66%, 98.74% |
| 2024 | Atlas et al. (88) | Maximally stable extremal regions approach, CNN, ELSTM, DPFE | DIARETDB2 | Specificity: 98.91%; sensitivity: 98.67% |
| 2023 | Li et al. (89) | CNN, ResNet-50, customized computer vision algorithm | Local dataset: University of California | Specificity: 74.5%; sensitivity: 100%; F1-score: 0.932; accuracy: 90.7% |
| 2023 | Alwakid et al. (90) | CLAHE filtering approach and ESRGAN network, Inception-V3 model | APTOS datasets | Case1: accuracy: 98.7%; case 2: accuracy: 80.87% |
| 2023 | Saranya et al. (91) | CNN, U-Net, morphological operations | Messidor, STARE DIARETDB1, IDRiD | Specificity: 99%; sensitivity: 89%; accuracy: 95.65% |
| 2023 | Kiliçarslan (92) | ResNet-50, YOLOv5, and VGG-19 | EyePACS | Accuracy: 93.38%, 94.75%, 91.72% |
| 2022 | Mondal et al. (93) | ResNeXt, DenseNet101, CLAHE | DIARETDB1, APTOS19 | Accuracy: 96.98%; precision: 0.97; recall: 0.97 |
| 2023 | Xia et al. (94) | MGNet, U-Net | IDRiD, DIARETDB1 | Accuracy: 99.12%; sensitivity: 53.53%; specificity: 99.66% |
| 2021 | Goel et al. (95) | DL, transfer learning, VGG16 | IDRiD | Accuracy: 91.8%; F1-score: 81.5%; precision: 81.5% |
| 2022 | Zhang et al. (96) | Inception-V3 model, class activation mapping | Kaggle | Sensitivity: 0.925; specificity: 0.907; harmonic mean: 0.916; AUC: 0.968 |
| 2020 | Hacisoftaoglu et al. (97) |
ResNet50, GoogLeNet, AlexNet models | Messidor-2, EyePACS, Messidor, IDRiD | Specificity: 99.1%; sensitivity: 98.2%; accuracy: 98.6% |
| 2020 | Li et al. (98) | DL system, inception ResNet-2 | Local dataset: ZOC | Sensitivity: 97.6%; specificity: 99.4%; AUC: 99.9% |
| 2019 | Chowdhury et al. (38) | Classifier: random forest and Naïve Bayes classifier approach | Messidor, Tele-Ophtha, DIARETDB1, DIARETDB0 | Accuracy: 93.58% |
| 2018 | Suriyal et al. (99) | DeepCNNs, MobileNets | Kaggle | Accuracy: 73.3% |
| 2017 | Gondal et al. (54) | CNNs, for generate the class activation | DIARETDB1 | AUC: 95.4% |
| 2016 | Paing et al. (55) | Artificial neural networks, and ROI localization | DIARETDB1 | Accuracy: 96% |
| 2015 | Guo et al. (100) | Multi-class discriminant analysis, wavelet transformation, and discrete cosine transformation | Real-world dataset | Accuracy: 90.9% |
APTOS, Asia Pacific Tele-Ophthalmology Society; AUC, area under the curve; CLAHE, contrast limited adaptive histogram equalization; CNN, convolutional neural network; DIARETDB, diabetic retinopathy database; DL, deep learning; DPFE, double-pierced feature extraction; ELSTM, enhanced long short-term memory; ESRGAN, enhanced super-resolution generative adversarial network; EyePACS, eye picture archive communication system; HRF, high-resolution fundus; IDRiD, Indian diabetic retinopathy image dataset; Messidor, methods to evaluate segmentation and indexing techniques in the field of retinal ophthalmology; MGNet, multi-scale gated network; ROI, region of interest; STARE, structured analysis of the retina; VGG, visual geometry group; YOLOv5, you only look once version 5; ZOC, Zhongshan Ophthalmic Center.