Feature Transformation Network based on Correlation Distribution Graph for Disease Diagnosis

Gene expression data play a crucial role in disease diagnosis. Transforming such non-image data into spatially structured image representations enables the ap-plication of convolutional neural networks (CNNs) to im-prove diagnostic accuracy. However, existing methods of-ten fail to capture complex spatial relationships among genes and overlook biologically meaningful interactions. Additionally, the spatial arrangement in generated images is usually fixed and unoptimized, limiting the model's capacity to learn effective spatial dependencies. To ad-dress these challenges, we propose a novel method, Feature Transformation Network based on Correlation Distribution Graph (FTNCDG), to convert high-dimensional gene expression data into image-like representations suitable for CNN-based classification. This model integrates gene screening and relation extraction to identify disease-relevant gene sets with shared biological functions. We further introduce a coordinate search algorithm based on a graph attention network to assign genes to 2D image coordinates with minimal overlap. We evaluated FTNCDG on four datasets (TCGA-BRCA, TCGA-LUAD, GSE96058, and GSE25066), achieving respective accuracies of 0.845, 0.913, 0.965, and 0.876; area under the receiver operating characteristic curve (AUC) values of 0.814, 0.876, 0.954, and 0.825; F1-scores of 0.730, 0.849, 0.911, and 0.633; and recall values of 0.751, 0.861, 0.931, and 0.724. These results demonstrate the superiority of FTNCDG-CNN over state-of-the-art methods, highlighting the importance of incorporating biological context into spatial encoding and establishing a new benchmark for future research in intelligent gene to-image modeling.