Background: Cardiovascular diseases (CVDs) are a leading cause of death worldwide. Deep learning methods have been widely used in the field of medical image analysis and have shown promising results in the diagnosis of CVDs. Methods: Experiments were performed on 12-lead electrocardiogram (ECG) databases collected by Chapman University and Shaoxing People’s Hospital. The ECG signal of each lead was converted into a scalogram image and an ECG grayscale image and used to fine-tune the pretrained ResNet-50 model of each lead. The ResNet-50 model was used as a base learner for the stacking ensemble method. Logistic regression, support vector machine, random forest, and XGBoost were used as a meta learner by combining the predictions of the base learner. The study introduced a method called multi-modal stacking ensemble, which involves training a meta learner through a stacking ensemble that combines predictions from two modalities: scalogram images and ECG grayscale images. Results: The multi-modal stacking ensemble with a combination of ResNet-50 and logistic regression achieved an AUC of 0.995, an accuracy of 93.97%, a sensitivity of 0.940, a precision of 0.937, and an F1-score of 0.936, which are higher than those of LSTM, BiLSTM, individual base learners, simple averaging ensemble, and single-modal stacking ensemble methods. Conclusion: The proposed multi-modal stacking ensemble approach showed effectiveness for diagnosing CVDs.

Pediatric thoracic diseases present significant health risks to children. While chest X-rays are commonly used for diagnosing thoracic diseases, interpreting pediatric images comes with unique challenges such as anatomical variations, developmental differences, and potential artifacts. Deep learning offers promise in addressing these challenges, yet its effectiveness is hindered by the limited availability of pediatric chest X-ray data. To overcome this limitation, we introduce the dual-masked autoencoders (dual-MAE) algorithm, consisting of online and target networks with encoder and decoder modules. These networks are optimized by minimizing three losses: between the reconstructed image of the online network and the target network, between the input image and the reconstructed image of the online network, and between the input image and the reconstructed image of the target network. To learn efficiently from pediatric chest X-rays, we employ a two-step training strategy: pretraining the dual-MAE model on adult chest X-rays, then fine-tuning it on pediatric X-rays for diagnosing multi-labeled pediatric thoracic diseases. The proposed model exhibited superior performance with the highest mean AUC score (0.752), surpassing the ResNet-34 (0.669) and ViT-S (0.645) trained from scratch. Additionally, the dual-MAE model outperformed the ResNet-34 (0.697) and ViT-S (0.638), both pretrained on the ImageNet dataset and then fine-tuned on pediatric chest X-rays. Despite being pretrained on a significantly smaller number of X-rays compared to the ImageNet dataset, our model demonstrated better performance. Furthermore, it outperformed the ResNet-34 (0.712), ViT-S (0.673), and vanilla MAE method (0.735), all pretrained on adult chest X-rays and fine-tuned on pediatric chest X-rays. Even with only 50% of labeled pediatric chest X-ray images, dual-MAE demonstrated comparable performance to that of the vanilla MAE method and outperformed ResNet-34 and ViT-S fine-tuned with 100% labeled pediatric chest X-ray images.

Background/Objectives: Retinal optical coherence tomography (OCT) is essential for diagnosing ocular diseases, yet developing high-performing multiclass classifiers remains challenging due to limited labeled data and the computational cost of self-supervised pretraining. This study aims to address these limitations by introducing a curriculum-based self-supervised framework to improve representation learning and reduce computational burden for OCT classification. Methods: Two ensemble strategies were developed using progressive masked autoencoder (MAE) pretraining. We refer to this curriculum-based MAE framework as CurriMAE (curriculum-based masked autoencoder). CurriMAE-Soup merges multiple curriculum-aware pretrained checkpoints using weight averaging, producing a single model for fine-tuning and inference. CurriMAE-Greedy selects top-performing fine-tuned models from different pretraining stages and ensembles their predictions. Both approaches rely on one curriculum-guided MAE pretraining run, avoiding repeated training with fixed masking ratios. Experiments were conducted on two publicly available retinal OCT datasets, the Kermany dataset for self-supervised pretraining and the OCTDL dataset for downstream evaluation. The OCTDL dataset comprises seven clinically relevant retinal classes, including normal retina, age-related macular degeneration (AMD), diabetic macular edema (DME), epiretinal membrane (ERM), retinal vein occlusion (RVO), retinal artery occlusion (RAO), and vitreomacular interface disease (VID) and the proposed methods were compared against standard MAE variants and supervised baselines including ResNet-34 and ViT-S. Results: Both CurriMAE methods outperformed standard MAE models and supervised baselines. CurriMAE-Greedy achieved the highest performance with an area under the receiver operating characteristic curve (AUC) of 0.995 and accuracy of 93.32%, while CurriMAE-Soup provided competitive accuracy with substantially lower inference complexity. Compared with MAE models trained at fixed masking ratios, the proposed methods improved accuracy while requiring fewer pretraining runs and reduced model storage for inference. Conclusions: The proposed curriculum-based self-supervised ensemble framework offers an effective and resource-efficient solution for multiclass retinal OCT classification. By integrating progressive masking with snapshot-based model fusion, CurriMAE methods provide high performance with reduced computational cost, supporting their potential for real-world ophthalmic imaging applications where labeled data and computational resources are limited.

Masked autoencoders (MAE) have emerged as a powerful framework for self-supervised learning by reconstructing masked input data. However, determining the optimal masking ratio requires extensive experimentation, resulting in significant computational overhead. To address this challenge, we propose CurriMAE, a curriculum-based training approach that progressively increases the masking ratio during pretraining to balance task complexity and computational efficiency. In CurriMAE, the training process spans 800 epochs, with the masking ratio gradually increasing in four stages: 60% for the first 200 epochs, followed by 70%, 80%, and finally 90% in the last 200 epochs. This progressive masking approach, inspired by curriculum learning, allows the model to learn from simpler tasks before tackling more challenging ones. To ensure stable convergence, a cyclic cosine learning rate scheduler is employed, resetting every 200 epochs, effectively dividing the training process into four distinct stages. At the end of each stage, corresponding to one complete cycle of the learning rate schedule, a snapshot model is saved, resulting in four pretrained models. These snapshots are then fine-tuned to obtain the final classification results. We evaluate CurriMAE on multi-labeled pediatric thoracic disease classification, pretraining the model on CheXpert and ChestX-ray14 datasets, and fine-tuning it on PediCXR. Experimental results show that CurriMAE outperforms ResNet, ViT-S, and standard MAE, achieving superior performance while reducing computational cost. These findings establish CurriMAE as an effective and scalable self-supervised learning framework for medical imaging applications.