Cancer is a leading cause of death globally, and analyzing digital pathology images for cancer diagnosis and treatment is a complex problem due to the high data volume, the large size of each image, lack of annotated data, and computational demands. To improve care and expand access, there is a need for fast, efficient and resource-constrained computation models accessible for all clinicians and researchers. Deep learning methods applied to tumour tissue images have improved detection, diagnosis and characterisation of cancer tumours in clinical routine. However, traditional machine learning models require annotated data and are limited in scope, while self-supervised foundation models demand vast datasets and consume extensive time and energy, creating accessibility barriers. Advanced artificial intelligence techniques like contrastive learning, data augmentation, data curation, and optimization strategies provide innovative approaches to achieve sample-efficient learning while maintaining high accuracy and generalizability. The overall aim of this project is to develop both methodologies and applications in AI-based precision pathology, with a focus on resource-constrained environments. By optimizing AI approaches for limited-resource settings, this project seeks to advance precision diagnostics and computer vision capabilities, ensuring broader access and impact. The project is based on in-house internationally unique, large and population representative studies, including more than five thousand of gigapixel Whole Slide Images (WSIs) and registry-based clinical information. Feasibility and potential impact on cancer and computer vision research is high. My hypothesis is that incorporating domain knowledge into deep learning models will lead to learn more effectively from smaller datasets by producing higher-level abstractions and filtering out noise.