This paper presents a deep learning system for computational pathology that uses weakly supervised learning to train classification models on whole slide images (WSIs) without the need for pixel-level annotations. The system leverages multiple instance learning (MIL) to classify tiles within a WSI based on slide-level diagnoses, which are readily available from pathology reports. The authors collected large datasets (44,732 slides) for prostate cancer, basal cell carcinoma (BCC), and breast cancer metastases, representing a significant improvement over existing datasets. The proposed method achieved high accuracy (AUCs > 0.98) across all cancer types, demonstrating its potential for clinical application. The system can help pathologists by excluding 65–75% of slides while maintaining 100% sensitivity, significantly reducing their workload. The authors also discuss the challenges of technical variability in slide preparation and scanning, and show that their model generalizes well to different scanners and slide preparation methods. The results highlight the feasibility of using weakly supervised learning to train accurate classification models at a large scale, paving the way for the deployment of computational decision support systems in clinical practice.This paper presents a deep learning system for computational pathology that uses weakly supervised learning to train classification models on whole slide images (WSIs) without the need for pixel-level annotations. The system leverages multiple instance learning (MIL) to classify tiles within a WSI based on slide-level diagnoses, which are readily available from pathology reports. The authors collected large datasets (44,732 slides) for prostate cancer, basal cell carcinoma (BCC), and breast cancer metastases, representing a significant improvement over existing datasets. The proposed method achieved high accuracy (AUCs > 0.98) across all cancer types, demonstrating its potential for clinical application. The system can help pathologists by excluding 65–75% of slides while maintaining 100% sensitivity, significantly reducing their workload. The authors also discuss the challenges of technical variability in slide preparation and scanning, and show that their model generalizes well to different scanners and slide preparation methods. The results highlight the feasibility of using weakly supervised learning to train accurate classification models at a large scale, paving the way for the deployment of computational decision support systems in clinical practice.