Skip to main content
. 2025 Apr 16;15(5):4816–4846. doi: 10.21037/qims-24-1791

Table 5. Exudates detection approaches.

Year Reference Methods Database Results with evaluation metrics
2024 Van Do et al. (63) Two-stage approach with CNN, VGGUnet, and RITM model DDR, IDRiD Mean Dice: 60.8%, 76.6%; mean IoU scores: 45.5%, 62.4%
2024 Guo et al. (64) Deep multi-scale model, FCN for lesion segmentation E-Ophtha, DDR, DIARETDB1, IDRiD DSC: 69.31%; AUPR: 68.84%
2024 Maiti et al. (65) Enriched encoder-decoder model with CLSTM and RES unit DIARETDB0, Messidor, DIARETDB1, IDRiD Overall accuracy: 97.7%; while accuracies of individual dataset respectively: 97.81%, 98.01%, 98.23%, 96.76%
2023 Farahat et al. (66) U-Net and YOLOv5, Leaky ReLU Cheikh Zaïd Foundation’s Ophthalmic Center Specificity: 85%; sensitivity: 85%; accuracy: 99%
2024 Naik et al. (67) DenseNet121 APTOS-2019 Accuracy: 96.64%
2023 Tang et al. (68) Patch-wise density loss, global segmentation loss, discriminative edge inspection IDRiD Precision: 84.03%; recall: 65.14%; F1-score: 68.54
2022 Reddy and
Gurrala (69)
Hybrid DLCNN-MGWO-VW, DSAM, and DDAM IDRiD Accuracy: 96.0%
2022 Hussain et al. (70) YOLOv5M, classification-extraction-superimposition mechanism EyePACS Accuracy: 100%
2023 Sangeethaa and
Jothimani (71)
Median filtering, CLAHE Local dataset: Aravind Eye Hospital, Coimbatore Accuracy: 95%
2021 Sudha and Ganeshbabu (72) Deep neural network, VGG-19, gradient descent, structure tensor Kaggle Sensitivity: 82%; accuracy: 96%
2021 Cincan et al. (73) GoogLeNet, SqueezeNet, and ResNet50 models SUSTech_SYSU Accuracy: 0.928%
2021 Kurilová et al. (74) Faster R-CNN, ResNet-50, region proposed network DIARETDB1, Messidor, E-Ophtha-EX AUC: 97.27%, 88.5%, 100%
2020 Pan et al. (75) ResNet50, VGG-16, DenseNet, multi-label classification 2nd Affiliated Hospital, Eye Center at Zhejiang University Hospital AUC: 96.53%
2020 Theera-Umpon et al. (76) Adaptive histogram equalization, hierarchical ANFIS, multilayer perceptron DIARETDB1 AUC: 99.8%
2020 Thulkar et al. (77) Tree-based forward search approach, SVM DIARETDB1, IDRiD, IDEDD Specificity: 97.1%; sensitivity: 100%
2019 Khojasteh et al. (58) DRBM, ResNet-50, CNNs E-Ophtha, DIARETDB1 Sensitivity: 99%; accuracy: 98%
2019 Chowdhury et al. (38) Random forest and Naïve Bayes classifier approach DIARETDB1, DIARETDB0 Accuracy: 93.58%
2018 Lamet al. (37) CNNs, AlexNet, GoogLeNet, VGG16, Inception-V3, ResNet E-Ophtha AUC: 95%
2017 Gondal et al. (54) CNNs, for generate the class activation mappings DIARETDB1 AUC: 95.4%
2016 Paing et al. (55) Artificial neural networks, and ROI iocalization DIARETDB1 Accuracy: 96%
2015 Prentašić and Lončarić (78) Deep CNNs DIARETDB1 PPV: 77%; sensitivity: 77%;
F1-score: 77%

ANFIS, adaptive network-based fuzzy inference system; APTOS, Asia Pacific Tele-Ophthalmology Society; AUC, area under the curve; AUPR, area under the precision-recall curve; CLAHE, contrast-limited adaptive histogram equalization; CLSTM, contextual long-short term memory; CNN, convolutional neural network; DDAM, disease-dependent attention module; DLCNN, deep learning convolutional neural network; DRBM, discriminative restricted Boltzmann machine; DSC, dice similarity coefficient; DSAM, disease-specific attention module; DDR, dataset for diabetic retinopathy; DIARETDB, diabetic retinopathy database; E-Ophtha, ophthalmology; EX, exudates; EyePACS, eye picture archive communication system; FCN, fully convolutional network; IDEDD, Indian diabetic eye diseases dataset; IDRiD, Indian diabetic retinopathy image dataset; IoU, intersection over union; Messidor, methods to evaluate segmentation and indexing techniques in the field of retinal ophthalmology; MGWO, modified grey-wolf optimizer; PPV, positive predictive value; R-CNN, region-based convolutional neural network; ReLU, rectified linear unit; RES, residual extended skip; RITM, reviving iterative training with mask guidance; ROI, region of interest; SUS, Southern University of Science and Technology; SVM, support vector machine; SYSU, Sun Yat-sen university; VGG, visual geometry group; VW, variable weight; YOLOv5, you only look once version 5; YOLOv5M, you only look once version 5 medium.