Researchers at the BC Cancer Registry developed a framework using Attention-Based Multiple Instance Learning (ABMIL) to train deep learning classifiers for tumor group classification without requiring per-report human annotation. The method leverages routine operational patient-level labels to distil a large, noisily-labeled corpus into a compact, high-quality per-report training dataset by using attention mechanisms to recover the link between patients and their pathology reports.
- The approach utilizes ABMIL to bridge the gap between patient-level labels and individual pathology reports.
- A classifier fine-tuned on the distilled dataset achieved a macro F1 of 0.83.
- This performance outperforms established baselines across most tumor groups.
- The framework automates cancer registry workflows without additional annotation or large-scale computing infrastructure.
By turning routine operational labels into high-quality training data, ABMIL offers a practical and accessible route to automating labor-intensive coding tasks in cancer registries.