MultiVIT: Cross‐Region Vision Transformers for Structured Regression in High‐Resolution Histopathology Images

Abstract

Continuous regression on high‐resolution images remains a challenging problem due to spatial heterogeneity, long‐range dependencies, and scale variability, particularly in digital histopathology. In this paper, we propose MultiVIT, a Vision Transformer framework designed for structured spatial regression from high‐resolution images. The proposed architecture introduces a cross‐region attention mechanism that explicitly models interactions between complementary spatial regions, enabling more informative representations than region‐independent approaches. Unlike discretized formulations, MultiVIT directly predicts continuous density values using a constrained regression head, improving stability and interpretability. Experiments conducted on a public immunohistochemistry dataset demonstrate that MultiVIT outperforms CNN and standard ViT baselines, achieving a reduction of MAE from 4.91 to 3.84 and improving R2$$ {R}^2 $$ from 0.73 to 0.84. Extended baselines (Swin‐B, TransMIL, HIPT) and cross‐dataset validation on TCGA‐CRC and Lizard further confirm generalizability. Ablation studies confirm the effectiveness of cross‐region attention and constrained regression. These results highlight MultiVIT as an effective and general framework for structured regression in high‐resolution image analysis, while also identifying key limitations related to dataset scale and generalization.