Abstract:
It is still very hard to combine molecular profiling with histopathological imaging in computational pathology. Reason bulk RNA sequencing only shows the average amount of gene expression and does not give you the spatial information you need to put tissue back together. This study investigates the feasibility of producing histopathological images directly from RNA-seq data using deep generative models, potentially enabling cost-effective imaging of molecular anomalies and the integration of genomic and morphological characterizations. We evaluated four generative architectures—Conditional GAN, Conditional VAE, Conditional Diffusion Model, and a novel Hybrid GAN-VAE—employing a dataset comprising 50 patients with paired RNA-seq profiles (16,383 genes) and H&Estained histopathology images (128×128 pixels). The dataset was divided at the patient level (70% for training, 15% for validation, and 15% for testing) to make sure that the generalization test was strong. We used RNA conditioning to train the models and then tested them with three different measures: the Structural Similarity Index (SSIM), the Learned Perceptual Image Patch Similarity (LPIPS), and the Fr´echet Inception Distance (FID). The Hybrid GAN-VAE had a better perceptual quality (LPIPS: 0.347), which was 16% better than the baseline GAN. The Conditional VAE, on the other hand, had the best statistical fidelity (FID: 275.64). Deep generative models can create realistic histopathological images from RNA-seq data; however, bulk sequencing compromises spatial information, resulting in decreased reconstruction accuracy. The Hybrid architecture’s superior perceptual performance demonstrates the efficacy of multi-objective training methodologies that incorporate both adversarial and reconstruction losses. Future research incorporating spatially-resolved transcriptomics could significantly enhance the quality of synthesis, facilitating its application in diagnostic procedures, medical education, and precision medicine research.
