One View for All: Unified Slide Representation in Pan-Cancer Diagnosis and Localization

Whole-slide imaging has transformed digital pathology, yet extracting clinically meaningful information from standard hematoxylin and eosin (H&E) slides remains challenging due to complex tissue heterogeneity. While AI holds great promise to assist with digital pathology, developing current AI models typically relies on exhaustive manual dataset annotations. This creates a training bottleneck that severely limits the scalability of AI models. But what if the exact training labels we need are already available? Instead of relying on cumbersome manual annotations, our model leverages the rich diagnostic information already documented in routine pathology reports as direct training labels. This approach completely bypasses the annotation bottleneck, enabling scalable computational pathology. We developed the AI model using a multi-institutional cohort of 10,359 H&E whole-slide images (WSIs) from the Cooperative Human Tissue Network (CHTN), utilizing standard pathology labels (tissue category, cancer type, anatomic site) as ground truth. Model performance was evaluated on three slide-level tasks: 4-category tissue classification, 14-class malignant cancer classification, and 3-category solid tumor group classification (carcinoma, sarcoma, melanoma). Tumor localization was validated against expert pathologist annotations on a subset of 300 WSIs spanning 14 cancer types. For tumor localization, the model achieved high spatial agreement with expert pathologist annotations across 14 diverse cancer types, yielding a mean Dice coefficient of 0.872. It also demonstrated strong performance across all classification tasks, with macro-AUCs of 97.6% (4-category tissue), 98.3% (14-class malignant cancer), and 99.4% (3-category solid tumor group). This study shows that routine pathology labels can support slide-level classification and tumor localization directly from H&E WSIs, offering a scalable and practical strategy for dig

Learning Objectives

1. Describe how routine pathology reports can replace exhaustive manual annotations to train scalable clinical AI models.
2. Assess the model's accuracy in pan-cancer slide-level classification and expert-level tumor localization from H&E WSIs.
3. Recognize how this scalable approach overcomes training bottlenecks to enable seamless clinical AI integration.