Automatic report-based assessment of radiology-pathology concordance in surgical patients using BERT and DPCNN

Assessing radiology-pathology concordance is important for retrospective audit, educational feedback, and quality assurance in radiology practice. However, automated concordance assessment remains challenging because of imbalanced data and the fuzzy semantic matching required between complex radiological and pathological diagnostic texts. This study introduces an automated deep learning framework to retrospectively assess radiology-pathology concordance in authentic clinical reports, serving as a focused proxy for one aspect of diagnostic accuracy rather than a comprehensive measure of overall radiology report quality. To overcome dataset imbalance, we applied targeted resampling techniques. We then integrated Bidirectional Encoder Representations from Transformers (BERT) with a Deep Pyramid Convolutional Neural Network (DPCNN) to accurately extract diagnostic semantics and match radiological conclusions against gold-standard pathological findings. Validated on a dataset of follow-up records from a Grade III Level A hospital, our proposed BERT-DPCNN model achieved an accuracy of 91.51%. These findings suggest that the proposed method can support automated retrospective identification of concordant and discordant radiology-pathology report pairs. Ultimately, this approach provides an automated tool with promising internal performance for retrospective auditing and educational feedback, while further external validation is required before broader clinical application.