Enhancing imbalanced dataset diagnosis using class-based input image composition

Abstract

Deep learning models often falter when faced with small, imbalanced datasets or degraded image quality, leading to unacceptably high false prediction rates. To bridge this gap, we introduce class-based image composition. This technique reformulates training inputs by fusing multiple intra-class images into unique composite input images (CoImg). By concentrating information density and amplifying intra-class variance, CoImg forces the model to discern subtle, nuanced disease patterns that might otherwise be lost. We validated this approach using the optical coherence tomography dataset for image-based deep learning methods (OCTDL), a collection of seven imbalanced retinal disease scan categories. From this, we engineered Co-OCTDL: a perfectly balanced variant where each training sample exists as a 3×1 layout composite. To measure the impact of this new representation, we benchmarked the original dataset against its composite counterpart using a VGG16 architecture. Precision was paramount. We maintained identical hyperparameters and model structures across all experiments to ensure a rigorous, fair comparison. The results were transformative. While baseline datasets struggled, the enhanced Co OCTDL achieved a near-perfect F1-score of 0.995 and an AUC of 0.9996. The method effectively neutralized the risks of class imbalance. It didn't just improve the numbers; it refined the diagnostic reliability of the model.