This study introduces a novel semi-supervised few-shot learning framework designed to automatically diagnose two critical retinal disorders: pathological myopia (PM) and myopic traction maculopathy (MTM). By employing a large-scale collection of 99,039 high-quality fundus and OCT images, the proposed framework effectively tackles the dual challenges of limited expert-annotated data and imbalanced class distributions. Central to our approach is the integration of Meta Pseudo Labels (MPL) with a robust multi-head attention mechanism, which jointly enhances feature extraction and classification accuracy under constrained supervision. For PM, a dedicated labeler model leverages both manually annotated and pseudo-labeled fundus images to achieve a test accuracy of 99.36%, while for MTM, a hierarchical classification model distinguishes six grades of disease severity (T0–T5) with an accuracy of 98.40%. Extensive experiments underscore the efficacy of rigorous data preprocessing procedures—including normalization, augmentation, and class-balanced loss adjustments—as well as ablation studies that validate the critical role of the attention module. In essence, our framework not only minimizes the dependency on large volumes of labeled data but also considerably reduces the annotation burden on clinicians. This paradigm shows promising potential for integration into community health screening programs, thereby facilitating early detection and more precise management of myopia-related complications.