Identify the most suitable deep regression approaches for WSI-based gene-expression prediction

Determine which deep regression approaches are most well-suited for predicting transcriptome-wide mRNA gene-expression profiles from hematoxylin and eosin-stained whole-slide images in computational pathology, given the extremely high-dimensional output space (N = 20,530 genes) and multiple viable modeling paradigms (e.g., direct multi-instance learning-based regression, patch-level direct regression, and contrastive learning-based methods).

Background

Gene-expression prediction from whole-slide images constitutes a high-dimensional regression task with numerous potential model designs spanning direct regression (with or without learnable aggregation) and contrastive representation alignment. The paper frames the landscape as having many plausible approaches and notes uncertainty about which approaches are most appropriate for this task.

This uncertainty motivates a comparative evaluation across several regression paradigms to provide guidance, but the overarching question of approach suitability in general settings is explicitly stated as unclear at the outset.