Named entity recognition in mammography radiology reports using a multilingual transfer learning approach

Breast cancer still represents an important challenge to human health all over the world. In this sense, early detection and accurate diagnosis are essential parameters for patient wel- fare. Mammography, being a major tool in this area, has resulted in an enormous amount of radiology reports that carry important diagnostic information. This large treasure, as produced by modern science, is often left unutilized due to the complexity of the text and subtleties of linguistic cues in it. This thesis aim to capitalize on the power BERT-based models bring in a bid to further increase the precision of entity recognition in these reports, being mindful of the dual challenge of fine-tuning medical text analysis and improving the diagnostic process. This thesis examines the performance of Bidirectional Encoder Representations from Transformers (BERT) based models in entity recognition of mammography radiology re- ports under precision, recall, and F1-score metrics. It involves various unsupervised, and supervised algorithms, exploring their advantages and limitations in handling medical text. The study first aims to maximize the learning potential by unfreezing all layers of the mod- els, which is theoretically beneficial. However, considering the evidence from early and rapid metric convergence, there might be overfitting problems. To address this, a more balanced approach of partial freezing of layers was experimented with, fine-tuning just the last two layers, which emerged as the best method that significantly reduced overfitting while enhancing generalization on unseen data. Indeed, a more detailed comparative analysis shows that entities referring to distinct and definite, well-characterized characteristics in the text, such as ”ASINTOMATICA” and ”BIRADS,” tend to consistently approach perfect scores. However, lower values of performance are presented by entities such as CONDUCTO DILATADO and ENTERA- MENTE ADIPOSO since they don’t occur very often and their resemblance to other terms is syntactic. This might even suggest that current BERT-based approaches are not that far from reaching the performance ceiling in this specific application, highlighted more by small performance discrepancies between the models. The best-performing models—Austin- MeDeBeRTa—have the edge on managing complex entities through specific architectural features and training regimes. Results indicate that, while BERT-based models indeed have a lot of potential for the automatic recognition of entities in clinical reports, their deployment must be fine-tuned to the specifics of the medical domain to stay robust, especially against syntax and sev- eral semantics-related issues. Future work should further refine these models to increase applicability and accuracy in a clinical setting.