Learning from Lost Provenance: Multiple Instance Learning for Cancer Registry Tumor Group Classification

Cancer registries need to extract tumor type information from pathology reports, traditionally relying on manual coding which is time-consuming and error-prone. Deep learning can automate this but requires report-level annotated data. In practice, registries only have patient-level labels (e.g., final diagnosis) that are not linked to individual reports. This paper introduces an Attention-Based Multiple Instance Learning (ABMIL) framework: it treats each patient's multiple reports as a bag, with the patient-level label as the bag label. The attention mechanism learns the contribution weight of each report to the final label, selecting the most informative reports. High-confidence report-label pairs are then distilled to create a clean, small dataset. Fine-tuning a classifier on this distilled dataset achieves macro F1 of 0.83 on BC Cancer Registry data, outperforming baselines. The method requires no extra expert annotation or large-scale GPU clusters—only routine operational outputs (patient labels and reports) are needed—offering a low-barrier AI upgrade path for cancer registries worldwide.