Img2ST-Net: efficient high-resolution spatial omics prediction from whole-slide histology images via fully...

Area of Science:

• Computational biology
• Artificial intelligence in genomics
• High-resolution spatial omics

Background:

• Spatial transcriptomics (ST) data generation is expensive and time-consuming.
• High-resolution ST platforms (e.g., Visium HD) present computational and modeling challenges.
• Existing methods struggle with the scale, sparsity, and low expression levels of high-resolution ST data.

Purpose of the Study:

• To develop an efficient and parallel framework for high-resolution spatial transcriptomics prediction from histology images.
• To address the limitations of conventional spot-by-spot regression methods for high-definition ST data.
• To improve the accuracy and spatial coherence of generated ST data.

Main Methods:

• Proposed Img2ST-Net, a fully convolutional network for dense, high-definition (HD) gene expression map generation.
• Modeled HD ST data as super-pixel representations, transforming the task into super-content image generation.
• Introduced SSIM-ST, a structural-similarity-based metric for evaluating high-resolution ST data robustness.

Main Results:

• Img2ST-Net outperformed state-of-the-art methods on Visium HD datasets (8 and 16 μm resolutions).
• Achieved high accuracy on Breast Cancer (MSE 0.1657) and Colorectal Cancer (MSE 0.7981) datasets.
• Demonstrated up to 28-fold acceleration in training time via region-wise modeling, with contrastive learning enhancing spatial fidelity.

Conclusions:

• Presented a scalable, biologically coherent framework for high-resolution ST prediction.
• Img2ST-Net provides an efficient and accurate solution for ST inference at scale.
• The framework lays groundwork for next-generation, resolution-aware ST modeling.