Retinitis pigmentosa (RP) is a group of inherited retinal disorders that lead to progressive photoreceptor degeneration and can culminate in blindness. Early identification of RP is clinically valuable for patient counseling and for enabling access to emerging therapies and monitoring, but early subjective symptoms are frequently absent. Routine color fundus images are widely obtained in clinical practice and can show RP-related signs such as peripheral pigmentary changes and vascular attenuation, yet these images are underused for large-scale screening because of limited specialist availability and the subtlety of early changes.
The study sought to determine whether deep learning models pretrained on general visual tasks could be finetuned to detect RP on color fundus photographs and to assess model interpretability. The work prioritized both diagnostic performance and explainability by combining transfer learning with visualization techniques to examine whether model attention aligned with clinically meaningful retinal regions.
The retrospective dataset comprised 321 color fundus images from 201 Japanese subjects at Keio University Hospital. Of these, 200 images came from 107 patients with clinically diagnosed RP and 121 images from 94 non-RP subjects; some subjects contributed multiple images. The study received approval from the Keio University School of Medicine Ethics Review Committee (Approval No. 20210089). Data access for the study was performed with protections in place; written informed consent was waived due to the retrospective design, and the authors did not have access to personally identifiable information during or after data handling.
The authors applied transfer learning by finetuning convolutional neural networks that were pretrained for general image-recognition tasks. The specific architectures evaluated were VGG16, ResNet50, and InceptionV3. This approach leverages pretrained feature extractors and adapts them to the RP classification task using the available fundus image dataset, which reflects routine clinical imaging where genetic or multimodal data may be absent.
Model performance was assessed using cross-validation procedures to evaluate robustness and generalizability of the finetuned networks. The paper emphasizes the use of cross-validation in contrast to some prior work that lacked such procedures. The study also incorporated explainability analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize model attention on fundus images.
Among the three finetuned architectures, InceptionV3 achieved the best classification accuracy, reported as 96.97%. The authors note that this level of accuracy is comparable to the average diagnostic accuracy of ophthalmologists, suggesting that finetuned deep learning models can reach clinically relevant performance for RP detection on color fundus images.
Grad-CAM visualizations indicated that the finetuned model tended to attend to retinal regions that are clinically meaningful for RP detection, including the peripheral retina and the posterior pole. The attention to these areas is consistent with expected RP-related findings such as peripheral degenerative changes and retinal vascular attenuation, supporting the potential interpretability of the model and its alignment with clinical reasoning.
The authors propose that an explainable deep-learning classifier for RP using routine color fundus images could function as a supportive screening tool. Such an aid may help identify subtle or early fundus abnormalities and triage patients for specialist evaluation, particularly in settings with limited ophthalmology resources. The study highlights that focusing on fundus photographs alone addresses practical screening scenarios where genetic testing or multimodal imaging is not routinely available.
The full image dataset is restricted from public deposition because of patient privacy and the approved institutional protocol. De-identified data may be made available to qualified researchers upon reasonable request to the Keio University School of Medicine Ethics Review Committee and subject to institutional review and data-use agreements. The authors reported that no specific funding supported the work and that no competing interests were declared.
Finetuning pretrained convolutional neural networks on a retrospective set of routine color fundus images produced a high-performing RP classifier, with InceptionV3 achieving 96.97% accuracy. Explainability analysis using Grad-CAM demonstrated model attention to clinically relevant retinal regions, bolstering interpretability. The results support the feasibility of deep-learning–based RP screening from fundus photography as a potential adjunct to ophthalmologic assessment, while data access and privacy constraints limit public dataset sharing.