The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.
This study is aimed at evaluating a deep transfer learning-based model for identifying diabetic retinopathy (DR) that was trained using a dataset with high variability and predominant type 2 diabetes (T2D) and comparing model performance with that in patients with type 1 diabetes (T1D). The Kaggle dataset, which is a publicly available dataset, was divided into training and testing Kaggle datasets. In the comparison dataset, we collected retinal fundus images of T1D patients at Chang Gung Memorial Hospital in Taiwan from 2013 to 2020, and the images were divided into training and testing T1D datasets. The model was developed using 4 different convolutional neural networks (Inception-V3, DenseNet-121, VGG1, and Xception). The model performance in predicting DR was evaluated using testing images from each dataset, and area under the curve (AUC), sensitivity, and specificity were calculated. The model trained using the Kaggle dataset had an average (range) AUC of 0.74 (0.03) and 0.87 (0.01) in the testing Kaggle and T1D datasets, respectively. The model trained using the T1D dataset had an AUC of 0.88 (0.03), which decreased to 0.57 (0.02) in the testing Kaggle dataset. Heatmaps showed that the model focused on retinal hemorrhage, vessels, and exudation to predict DR. In wrong prediction images, artifacts and low-image quality affected model performance. The model developed with the high variability and T2D predominant dataset could be applied to T1D patients. Dataset homogeneity could affect the performance, trainability, and generalization of the model.
Diabetes Dataset Kaggle Download
Download Zip 🔥 https://bytlly.com/2yGc24 🔥
Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Medical artificial intelligence (AI) has achieved significant progress in recent years with the notable evolution of deep learning techniques1,3,4. For instance, deep neural networks have matched or surpassed the accuracy of clinical experts in various applications5, such as referral recommendations for sight-threatening retinal diseases6 and pathology detection in chest X-ray images7. These models are typically developed using large volumes of high-quality labels, which requires expert assessment and laborious workload1,2. However, the scarcity of experts with domain knowledge cannot meet such an exhaustive requirement, leaving vast amounts of medical data unlabelled and unexploited.
Figure 1 gives an overview of the construction and application of RETFound. For construction of RETFound, we curated 904,170 CFP in which 90.2% of images came from MEH-MIDAS and 9.8% from Kaggle EyePACS33, and 736,442 OCT in which 85.2% of them came from MEH-MIDAS and 14.8% from ref. 34. MEH-MIDAS is a retrospective dataset that includes the complete ocular imaging records of 37,401 patients with diabetes who were seen at Moorfields Eye Hospital between January 2000 and March 2022. After self-supervised pretraining on these retinal images, we evaluated the performance and generalizability of RETFound in adapting to diverse ocular and oculomic tasks. We selected publicly available datasets for the tasks of ocular disease diagnosis. Details are listed in Supplementary Table 1. For the tasks of ocular disease prognosis and systemic disease prediction, we used a cohort from the Moorfields AlzEye study (MEH-AlzEye) that links ophthalmic data of 353,157 patients, who attended Moorfields Eye Hospital between 2008 and 2018, with systemic disease data from hospital admissions across the whole of England35. We also used UK Biobank36 for external evaluation in predicting systemic diseases. The validation datasets used for ocular disease diagnosis are sourced from several countries, whereas systemic disease prediction was solely validated on UK datasets due to limited availability of this type of longitudinal data. Our assessment of generalizability for systemic disease prediction was therefore based on many tasks and datasets, but did not extend to vastly different geographical settings. Details of the clinical datasets are listed in Supplementary Table 2 (data selection is introduced in the Methods section).
Stage one constructs RETFound by means of SSL, using CFP and OCT from MEH-MIDAS and public datasets. Stage two adapts RETFound to downstream tasks by means of supervised learning for internal and external evaluation.
a, Internal evaluation. Models are adapted to curated datasets from MEH-AlzEye by fine-tuning and internally evaluated on hold-out test data. b, External evaluation. Models are fine-tuned on MEH-AlzEye and externally evaluated on the UK Biobank. Data for internal and external evaluation are described in Supplementary Table 2. Although the overall performances are not high due to the difficulty of tasks, RETFound achieved significantly higher AUROC in all internal evaluations and most external evaluations. For each task, we trained the model with five different random seeds, determining the shuffling of training data, and evaluated the models on the test set to get five replicas. We derived the statistics with the five replicas. The error bars show 95% CI and the bar centre represents the mean value of the AUROC. We compare the performance of RETFound with the most competitive comparison model to check whether statistically significant differences exist. P value is calculated with the two-sided t-test and listed in the figure.
Label efficiency refers to the amount of training data and labels required to achieve a target performance level for a given downstream task, which indicates the annotation workload for medical experts. RETFound showed superior label efficiency across various tasks (Fig. 4). For heart failure prediction, RETFound outperformed the other pretraining strategies using only 10% of labelled training data, demonstrating the potential of this approach in alleviating data shortages. RETFound similarly showed superior label efficiency for diabetic retinopathy classification and myocardial infarction prediction. Furthermore, RETFound showed consistently high adaptation efficiency (Extended Data Fig. 4), suggesting that RETFound required less time in adapting to downstream tasks. For example, RETFound can potentially save about 80% of the training time required to achieve convergence for the task of predicting myocardial infarction, leading to significant reductions in computational costs (for example, credits on Google Cloud Platform) when appropriate mechanisms such as early stopping are used.
Label efficiency measures the performance with different fractions of training data to understand the amount of data required to achieve a target performance level. The dashed grey lines highlight the difference in training data between RETFound and the most competitive comparison model. RETFound performs better than the comparison groups with 10% of training data in 3-year incidence prediction of heart failure and myocardial infarction with modality of CFP and comparable to other groups with 45% of data in diabetic retinopathy MESSIDOR-2 and 50% of data on IDRID. The 95% CI of AUROC are plotted in colour bands and the centre points of the bands indicate the mean value of AUROC.
We explored the performance of different SSL strategies, that is, generative SSL (for example, masked autoencoder) and contrastive SSL (for example, SimCLR, SwAV, DINO and MoCo-v3), in the RETFound framework. As shown in Fig. 5, RETFound with different contrastive SSL strategies showed decent performance in downstream tasks. For instance, RETFound with DINO achieved AUROC of 0.866 (95% CI 0.864, 0.869) and 0.728 (95% CI 0.725, 0.731), respectively, on wet-AMD prognosis (Extended Data Fig. 5) and ischaemic stroke prediction (Fig. 5), outperforming the baseline SL-ImageNet (Supplementary Tables 3 and 4). This demonstrates the effectiveness of RETFound framework with diverse SSL strategies. Among these SSL strategies, the masked autoencoder (primary SSL strategy for RETFound) performed significantly better than the contrastive learning approaches in most disease detection tasks (Fig. 5 and Extended Data Fig. 5). All quantitative results are listed in Supplementary Table 4.
We show AUROC of predicting diabetic retinopathy, ischaemic stroke and heart failure by the models pretrained with different SSL strategies, including the masked autoencoder (MAE), SwAV, SimCLR, MoCo-v3 and DINO. The data for systemic disease tasks come from the MEH-AlzEye dataset. RETFound with MAE achieved significantly higher AUROC in most tasks. The corresponding quantitative results for the contrastive SSL approaches are listed in Supplementary Table 4. For each task, we trained the model with five different random seeds, determining the shuffling of training data, and evaluated the models on the test set to get five replicas. We derived the statistics with the five replicas. The error bars show 95% CI and the bar centre represents the mean value of the AUPR. We compare the performance of RETFound with the most competitive comparison model to check whether statistically significant differences exist. P value is calculated with the two-sided t-test and listed in the figure.
To gain insights into the inner-workings of RETFound leading to its superior performance and label efficiency in downstream tasks, we performed qualitative analyses of the pretext task used for self-supervised pretraining and task-specific decisions of RETFound (Extended Data Fig. 6). The pretext task of RETFound allows models to learn retina-specific context, including anatomical structures and disease lesions. As shown in Extended Data Fig. 6a, RETFound was able to reconstruct major anatomical structures, including the optic nerve and large vessels on CFP, and the nerve fibre layer and retinal pigment epithelium on OCT, despite 75% of the retinal image being masked. This demonstrates that RETFound has learned to identify and infer the representation of disease-related areas by means of SSL, which contributes to performance and label efficiency in downstream tasks. On top of the reconstruction-based interpretation, we further used an advanced explanation tool (RELPROP42) to visualize the salient regions of images conducive to classifications made by fine-tuned models in downstream tasks (Extended Data Fig. 6b). For ocular disease diagnosis, well-defined pathologies were identified and used for classification, such as hard exudates and haemorrhage for diabetic retinopathy and parapapillary atrophy for glaucoma. For oculomic tasks, we observed that anatomical structures associated with systemic conditions, such as the optic nerve on CFP and nerve fibre layer and ganglion cell layer on OCT, were highlighted as areas that contributed to the incidence prediction of systemic diseases (Extended Data Fig. 6b). 152ee80cbc
international code of signals pdf free download
how to download apps on older samsung smart tv