A multi-modal automated retinal eye disease classification using the ODIR-5K dataset was implemented

The gap

Evidence profile

Sourced from the future work of the source papers, classified as general, spanning 3 journals.

Supporting evidence — 3 representative gaps

• GDT-SwinKid: A hybrid model for precise renal lesion analysis (2026) · PLOS One · doi

  GDT-SwinKid is the latest advancement in the development of automated analytical capabilities for the automated analysis of kidney disease using deep learning. This is achieved by merging GDT, an extraction system based on Swin Transfor- mations (and the extraction of the primary and secondary structures) into a single high-quality backbone (the U-Net) to create a single effective model that not only achieves very high accuracy for segmentation (Dice: 0.95) and classification (AUC = 0.991), but also reduces parameter count and computational requirements compared to most current methods. When compared with other existing architectures, the current work consistently produced superior performance relative to the other transformer-based approaches or the current efficient CNN designs specifically, it provided enhanced accuracy in localizing lesions and showed greater consistency across a wide range of classes or categories of clinical diseases (cyst, tumor, stone) PLOS One | https://doi.org/10.1371/journal.pone.0349285 May 20, 2026 27 / 30 when performed across these multiple classes or categories. Furthermore, the results from our ablation studies showed that every one of the innovative design features contributes significantly to the overall success of the model. The visual evidence generated through the model’s interpretability (attention and Grad-CAM) supports the accuracy of predictions, leading to a high degree of confidence in the use of deep learning methods. Collectively, this paper provides a ground-breaking new high-performance model that surpasses all previous results in segmentation or classification of renal disease. 9.1. Future work will focus on Here are future works directions, presented in clear point wise, 1. Can expand the GDT-SwinKid framework to include multi-modal and multi-view imaging data, such as MRI and ultra- sound, aiming to support more comprehensive and realistic clinical scenarios. 2. Can integrate clinical metadata (e.g., patient demographics and lab results) alongside imaging to enhance diagnostic precision and boost clinical relevance. 3. Focus on rare lesion types and borderline cases by developing advanced augmentation methods or adopting few-shot and transfer learning strategies.

  generalfuture work
  Keywords: high model clinical learning accuracy current swinkid automated disease deep extraction based single segmentation classification
• Ultra-widefield optical coherence tomography angiography in diabetic retinopathy: from retinal lesions to choroidal metrics (2026) · Frontiers in Medicine · doi

  Several future research areas for UWF-OCTA in DR require atten- tion. Although UWF-OCTA has shown certain application value in DR assessment, current research and clinical practice still have numer- ous unresolved limitations, such as non-unified imaging specifica- tions, insufficient automatic lesion recognition accuracy, and unclear guidance value of quantitative indicators. In view of these practical bottlenecks, subsequent research of UWF-OCTA in DR can be pro- moted step by step in combination with short-term, medium-term and long-term layouts, so as to gradually improve its clinical applica- tion system. In the short term, deep learning can be prioritized to optimize the automatic lesion detection capability of UWF-OCTA. Flat NV is often difficult to identify accurately due to ILM segmentation limita- tions, peripheral scan images have relatively low signal-to-noise ratio, and severe DME will also disturb the integrity of retinal structure to a certain extent. Constructing deep learning models with diverse training data covering the above complex scenarios and conducting rigorous cross-device and cross-center verification will help to steadily improve the accuracy of automatic identification of key lesions. In the medium term, large-scale longitudinal and horizontal comparative studies should be carried out simultaneously to clar- ify the clinical value of UWF-OCTA-related indicators. Long-term follow-up of patients during the transition from NPDR to PDR and treatment can help explore the temporal correlation between choroidal and retinal vascular changes, and determine the predic- tive effect of relevant indicators on clinical outcomes. Meanwhile, prospective comparative studies between UWF-OCTA-based evaluation algorithms and FA or clinical examination are needed to confirm whether quantitative indicators such as NV area and vessel density can guide the formulation of clinical strategies including anti-VEGF treatment timing and panretinal photoco- agulation (PRP). In addition, further exploration of choroidal metrics is also an important part of medium-term research, including clarifying the sequence of choroidal thinning and reti- nal damage, verifying the response of choroidal parameters to anti-VEGF therapy, and standardizing the quantification method of choroidal flow voids, so as to establish choroidal metrics as effective clinical biomarkers. In the long run, establishing multi-center consensus and a standardized UWF-OCTA imaging system is the core task. Current inconsistencies in devices, scan protocols, segmentation algorithms and analysis methods limit the generalizability of research results. Developing unified standards for imaging acqui- sition, reporting and quality control through multi-center coop- eration will lay the foundation for cross-study data integration and promote the standardized application of UWF-OCTA in DR clinical management.

  generalfuture work
  Keywords: octa clinical term choroidal indicators value imaging automatic medium long cross center tion certain application
• SSLVNet: an explainable multi-modal hybrid framework integrating self-supervised learning and vision transformers for multi-label classification of retinal diseases (2026) · Frontiers in Artificial Intelligence · doi

  A multi-modal automated retinal eye disease classification using the ODIR-5K dataset was implemented. The model com- bines self-supervised representation learning with hybrid deep neural architectures, including CNNs and ViT, to efficiently cap- ture both localized retinal abnormalities and wider structural designs. Additionally, attention-based mechanisms and graph- oriented feature modeling are incorporated to enhance the under- standing of complex lesion relationships within retinal images. Experimental results indicate the proposed system delivers reliable performance across multiple retinal disease categories. Moreover, interpretability is improved through Grad-CAM visualizations that highlight influential retinal regions, while uncertainty estima- tion increases the confidence and trustworthiness of predictions in clinical use. Further research may also focus on developing lightweight transformer architectures that reduce computational cost while maintaining accuracy. In addition, deploying the model in real-time clinical screening environments and integrating it into hospital decision-support systems would improve practical usability and assist clinicians in the early detection of retinal diseases.

  generalfuture work
  Keywords: retinal disease model architectures clinical multi modal automated classification using odir dataset implemented bines self

Questions about this gap