Domain adaptation aims to reduce the distribution gap between the training data (source domain) and the target data. This enables effective predictions even for domains not seen during training. However, most conventional domain adaptation methods assume a single source domain, making them less suitable for modern deep learning settings that rely on diverse and large-scale datasets. To address this limitation, recent research has focused on Multi-Source Domain Adaptation (MSDA), which aims to learn effectively from multiple source domains. In this paper, we propose Efficient Domain Transition for Multi-source (EDTM), a novel and efficient framework designed to tackle two major challenges in existing MSDA approaches: (1) integrating knowledge across different source domains and (2) aligning label distributions between source and target domains. EDTM leverages an ensemble-based classifier expert mechanism to enhance the contribution of source domains that are more similar to the target domain. To further stabilize the learning process and improve performance, we incorporate imitation learning into the training of the target model. In addition, Maximum Classifier Discrepancy (MCD) is employed to align class-wise label distributions between the source and target domains. Experiments were conducted using Digits-Five, one of the most representative benchmark datasets for MSDA. The results show that EDTM consistently outperforms existing methods in terms of average classification accuracy. Notably, EDTM achieved significantly higher performance on target domains such as Modified National Institute of Standards and Technolog with blended background images(MNIST-M) and Street View House Numbers(SVHN) datasets, demonstrating enhanced generalization compared to baseline approaches. Furthermore, an ablation study analyzing the contribution of each loss component validated the effectiveness of the framework, highlighting the importance of each module in achieving optimal performance.

Early detection of diabetic retinopathy (DR), media haze (MH), optic disc cupping (ODC), and glaucoma is crucial for preventing vision loss. However, timely diagnosis is often constrained by limited specialist availability and high diagnostic costs. This study proposes a You Only Look Once (YOLO)-based deep learning (DL) framework for the automated classification of fundus images into disease-specific categories. We unified diverse annotations from the Retinal Fundus Multi-Disease image Dataset (RFMiD), RFMiD2.0, and the DR Fundus Image Dataset (DR-FID) by standardizing annotation files and class labels. A custom filtering module was used to isolate single-pathology cases, and dataset issues such as missing or corrupted files were identified and resolved. To handle class imbalance, we applied oversampling and undersampling methods. The dataset was re-engineered for lightweight, accurate classification with YOLOv11, utilizing offline preprocessing tailored for retinal images. The dataset design leverages YOLOv11’s multi-class classification framework to achieve high performance on resource-constrained devices. This tailored approach outperforms preparing datasets solely through cloud-based platforms like Roboflow. The proposed model uses a lightweight YOLOv11 architecture, resulting in faster inference and lower memory requirements than conventional Convolutional Neural Networks (CNNs), such as Residual Networks (ResNets) or Visual Geometry Group (VGG) networks. Delivering high accuracy with minimal resource use, the model shows no signs of divergence or overfitting. Confusion matrices and class-wise metrics confirm consistent performance. The proposed framework achieves improved performance, with 94.78% accuracy, 96.12% specificity, 79.61% precision, 83.61% recall, and an 81.14% F1-score, demonstrating strong generalization to the internal held-out test set.

Lightweight deep learning models are increasingly required in resource-constrained environments such as mobile devices and the Internet of Medical Things (IoMT). Multi-head convolution with channel attention can facilitate learning activations relevant to different kernel sizes within a multi-head convolutional layer. Therefore, this study investigates the capability of novel lightweight models incorporating residual multi-head convolution with channel attention (ResMHCNN) blocks to classify medical images. We introduced three novel lightweight deep learning models (BT-Net, LCC-Net, and BC-Net) utilizing the ResMHCNN block as their backbone. These models were cross-validated and tested on three publicly available medical image datasets: a brain tumor dataset from Figshare consisting of T1-weighted magnetic resonance imaging slices of meningioma, glioma, and pituitary tumors; the LC25000 dataset, which includes microscopic images of lung and colon cancers; and the BreaKHis dataset, containing benign and malignant breast microscopic images. The lightweight models achieved accuracies of 96.9% for 3-class brain tumor classification using BT-Net, and 99.7% for 5-class lung and colon cancer classification using LCC-Net. For 2-class breast cancer classification, BC-Net achieved an accuracy of 96.7%. The parameter counts for the proposed lightweight models—LCC-Net, BC-Net, and BT-Net—are 0.528, 0.226, and 1.154 million, respectively. The presented lightweight models, featuring ResMHCNN blocks, may be effectively employed for accurate medical image classification. In the future, these models might be tested for viability in resource-constrained systems such as mobile devices and IoMT platforms.