Inter-document coreference resolution of abnormal findings in radiology documents

Area of Science:

• Medical Informatics
• Radiology
• Artificial Intelligence

Background:

• Tracking patient conditions over time is crucial for effective clinical management and treatment evaluation.
• Automated systems can enhance the monitoring of pertinent findings, aiding in condition management and assessing treatment outcomes.
• Lung cancer patients undergo serial computed tomography (CT) examinations, generating extensive data that requires careful analysis.

Purpose of the Study:

• To address the challenge of correlating lung mass observations across different documents from serial CT examinations in lung cancer patients.
• To develop an automated method for inter-document coreference resolution of radiological findings.

Main Methods:

• A probabilistic model was developed to quantify the likelihood that two findings from separate documents refer to the same entity.
• A greedy algorithm was employed, utilizing the probabilistic model to establish coreference links between identified findings.
• The methodology focuses on correlating observations of lung masses specifically.

Main Results:

• Preliminary evaluation demonstrated a precision of 72% for the inter-document coreference resolution task.
• The system achieved a recall of 63% in correlating findings across different CT examination documents.
• The probabilistic model and greedy algorithm showed promising performance in linking related radiological observations.

Conclusions:

• The presented methodology offers a viable approach for automated inter-document coreference resolution of radiological findings.
• This technique can significantly aid in tracking the progression of conditions like lung cancer by correlating findings over time.
• Improved tracking of findings can lead to better patient management and more accurate assessment of treatment efficacy.