Multi-Encode Fusion Architecture for Medical Image
Description
The effectiveness of medical image analysis systems is highly dependent on the quality, diversity, and representativeness of the dataset. In this study, a carefully curated multi-class medical dataset consisting of 20,000 images was utilized, encompassing multiple diagnostic categories including vitamin deficiencies and retinal blood vessel analysis. The dataset was designed to capture real-world variability in clinical imaging conditions, including differences in illumination, acquisition devices, color distributions, and anatomical structures. The dataset includes six primary classes: Vitamin A Deficiency, Vitamin B Deficiency, Vitamin C Deficiency, Vitamin D Deficiency, Vitamin E Deficiency, and Retina Blood Vessel Analysis. Each class exhibits distinct visual characteristics. Vitamin deficiency classes primarily differ in subtle texture patterns, color distributions, and surface irregularities, while retinal blood vessel images contain complex vascular structures with varying thickness, branching patterns, and contrast levels. These variations make the classification task inherently challenging and suitable for evaluating multi-domain feature learning. To ensure robust evaluation and avoid data leakage, the dataset is divided into training, validation, and test sets using a stratified sampling strategy. Approximately 70% of the data is allocated for training, while 15% is used for validation and 15% for independent testing. This distribution maintains class balance across all subsets and provides a reliable estimate of model generalization performance. In addition to image data, a structured clinical dataset containing 2,190 samples is incorporated to support multi-modal learning. This dataset includes multiple numerical features associated with diagnostic labels, enabling the model to capture complementary non- visual information.