Automated extraction of labels from abdominal radiology reports holds significant promise for enhancing both the accuracy and efficiency of neonatal necrotizing enterocolitis (NEC) diagnosis. Despite advancements in applying natural language processing (NLP) techniques to radiology reports, current efforts have primarily focused on chest radiology reports. Moreover, the lack of a publicly available abdominal X-ray dataset has limited extensive research into NEC diagnosis using abdominal X-rays (AXR). To address these challenges, we will utilize pediatric abdominal X-ray reports from a large medical institution and introduce AXpert. AXpert comes in two versions: one based solely on instruction fine-tuned large language models (LLMs), and another alternative version is a BERT-based model distilled from these fine-tuned models for improved inference and fine-tuning efficiency. AXpert can perform two tasks: (1) detecting the presence of NEC in children and (2) classifying the subtype of NEC into pneumatosis, portal venous gas, and free air. Comprehensive experimentation shows that AXpert outperforms baseline BERT models across all evaluation metrics. Furthermore, various factors affecting the performance of LLMs are examined, and the best combinations are identified based on the results. Additionally, the knowledge-distilled BERT model achieves performance on par with the LLM labelers and surpasses the performance of expert-annotated trained baseline BERT models. Our findings highlight AXpert's potential to reduce human labeling costs while maintaining high labeling accuracy. Moreover, AXpert provides precise image labeling, thereby advancing the automation of NEC diagnosis using AXR.

This presentation will be held in 2036 Palmer Commons. There will also be a remote viewing option via Zoom.