Arfa Jabin, Jyoti Singh Kirar, Shandar Ahmad
Whole Slide Imaging (WSI) has transformed digital pathology by enabling digitization of histological slides at gigapixel resolution. Artificial intelligence (AI), particularly deep learning (DL), models have leveraged WSIs to improve cancer diagnostics and, more recently, to infer transcriptomic profiles directly from tissue morphology. Although image analysis methods have made rapid progress and been widely reviewed before, specific challenges arise in dealing with domain-specific issues in WSI data. This review systematically surveys recent progress in AI-based analysis, focusing on two key predictive tasks namely, (i) cancer diagnosis and characterization and (ii) transcriptome or gene expression profile prediction from WSI samples. Giving a brief overview of WSI data collection and analysis in historical perspective, we review the techniques developed for diagnostic tasks, in the areas of model architectures such as Convolutional Neural Networks (CNNs), Vision Transformers, and Multiple Instance Learning (MIL). For the less studied but emerging problem of transcriptomic prediction, we discuss leading models such as HE2RNA, SEQUOIA, and tRNAsformer, their generalizability, and evaluation metrics (e.g. AUC, C-index, Pearson correlation). We highlight how AI models have achieved high diagnostic accuracy through specific, pan-cancer and multimodal foundation models and how the field is likely to evolve. AI-enabled WSI analysis shows substantial promise for both diagnostic and molecular inference tasks in oncology. However, key limitations such as spatial resolution constraints, lack of external validation, biological interpretability and clinical integration barriers need further development of methodologies in both directions. Specifically, future efforts need to address multimodal data integration, improved interpretability, and rigorous validation in diverse cohorts to realise the full potential of their clinical translation.