A deep learning system using weakly supervised learning was developed to classify whole slide images (WSIs) in pathology without requiring pixel-level annotations. The system was tested on 44,732 WSIs from 15,187 patients, achieving areas under the curve (AUC) above 0.98 for prostate cancer, basal cell carcinoma, and breast cancer metastases. This approach allows pathologists to exclude 65–75% of slides while maintaining 100% sensitivity. The system leverages the multiple instance learning (MIL) framework, where slide-level diagnoses are used as labels, enabling training on large-scale datasets without manual annotations. The model was evaluated on three large datasets: prostate core biopsies, skin lesions, and breast metastases to axillary lymph nodes. The results show that the system can accurately classify tumors in various tissue types, with AUCs of 0.986, 0.986, and 0.965 for prostate, BCC, and breast metastases, respectively. The system was also tested on different scanners and slide preparation methods, showing robustness to technical variability. The model's performance was compared to fully supervised learning, demonstrating that weakly supervised approaches can achieve clinical-grade accuracy without the need for extensive data curation. The study highlights the potential of computational pathology to improve diagnostic efficiency and accuracy in clinical practice.A deep learning system using weakly supervised learning was developed to classify whole slide images (WSIs) in pathology without requiring pixel-level annotations. The system was tested on 44,732 WSIs from 15,187 patients, achieving areas under the curve (AUC) above 0.98 for prostate cancer, basal cell carcinoma, and breast cancer metastases. This approach allows pathologists to exclude 65–75% of slides while maintaining 100% sensitivity. The system leverages the multiple instance learning (MIL) framework, where slide-level diagnoses are used as labels, enabling training on large-scale datasets without manual annotations. The model was evaluated on three large datasets: prostate core biopsies, skin lesions, and breast metastases to axillary lymph nodes. The results show that the system can accurately classify tumors in various tissue types, with AUCs of 0.986, 0.986, and 0.965 for prostate, BCC, and breast metastases, respectively. The system was also tested on different scanners and slide preparation methods, showing robustness to technical variability. The model's performance was compared to fully supervised learning, demonstrating that weakly supervised approaches can achieve clinical-grade accuracy without the need for extensive data curation. The study highlights the potential of computational pathology to improve diagnostic efficiency and accuracy in clinical practice.