Building blocks for complex tasks: Robust generative event extraction for radiology reports under domain shifts
This work addresses the challenge of domain adaptation in radiology report processing for clinical applications, offering incremental improvements in generalization and cost reduction.
The paper tackles the problem of extracting information from radiology reports that generalizes across exam modalities to reduce annotation needs, showing that multi-pass T5-based generative models outperform BERT-based classification approaches. It introduces a generative technique that decomposes complex tasks into subtask blocks and uses target-domain contexts during inference, improving single-pass models and enabling smaller models for clinical applications.
This paper explores methods for extracting information from radiology reports that generalize across exam modalities to reduce requirements for annotated data. We demonstrate that multi-pass T5-based text-to-text generative models exhibit better generalization across exam modalities compared to approaches that employ BERT-based task-specific classification layers. We then develop methods that reduce the inference cost of the model, making large-scale corpus processing more feasible for clinical applications. Specifically, we introduce a generative technique that decomposes complex tasks into smaller subtask blocks, which improves a single-pass model when combined with multitask training. In addition, we leverage target-domain contexts during inference to enhance domain adaptation, enabling use of smaller models. Analyses offer insights into the benefits of different cost reduction strategies.