Back to the list
Congress: ECR25
Poster Number: C-17749
Type: Poster: EPOS Radiologist (scientific)
Authorblock: R. Senkeev, M. Balbi, A. Veltri; Turin/IT
Disclosures:
Rouslan Senkeev: Nothing to disclose
Maurizio Balbi: Nothing to disclose
Andrea Veltri: Nothing to disclose
Keywords: Artificial Intelligence, Lung, Thorax, CT, Computer Applications-General, Patterns of Care
Purpose Millions of radiological examinations are conducted annually in developed countries, accumulating large repositories of digital health data and diagnostic images. These data, stored in Electronic Medical Records (EMR) and Picture Archive and Communication Systems (PACS) systems, represent a significant resource for medical research and clinical practice. However, this potential is largely underutilized due to the limited adoption of standardized structured reporting in radiology and the challenges associated with retrieving categorized information from documents written in natural, non-standardized language.The development of automatic categorization...
Read more Methods and materials To train the model, a database consisting of 10,840 unique sentences was extracted from all chest CT reports performed in 2021 (totalling 2001 reports) at the Radiology Service of the San Luigi Gonzaga University Teaching Hospital. Reports of CT studies that included other body parts, such as the abdomen and brain, were excluded.These sentences were manually annotated with labels across three main categories: findings, anatomical sites, and study descriptors. The dataset was split into training and testing sets, and a BERT model...
Read more Results The model’s performance was evaluated over 10 epochs, with a consistent decrease in both training and evaluation loss observed throughout the training process. Key metrics such as F1-score, ROC AUC, and accuracy were used to assess performance. The model demonstrated strong classification performance by the eighth epoch, achieving an F1-score of 0.980, an ROC AUC of 0.984, and an accuracy of 0.881. The stable decline in both training and evaluation loss indicated that the model was learning effectively and in the definitive...
Read more Conclusion The study demonstrates the possibility of adopting a Natural Language Processing (NLP) model, pre-trained on Italian language data and openly available, for use in a local offline framework for the classification of chest CT reports written in the Italian language. [fig 1] This work underscores the feasibility of integrating such an AI-based system into radiology, particularly in a context like Italian healthcare, where unstructured reporting is common. Furthermore, the continued development of similar models could significantly enhance the value of textual data and,...
Read more References Mun SK, Wong KH, Lo SB, Li Y, Bayarsaikhan S. Artificial Intelligence for the Future Radiology Diagnostic Service. Front Mol Biosci. 2021 Jan 28;7:614258.  Kim S, Lee CK, Kim SS. Large Language Models: A Guide for Radiologists. Korean J Radiol. 2024 Feb;25(2):126-133.  Devlin J, Chang M, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. North American Chapter of the Association for Computational Linguistics; 2019. Bressem KK, Adams LC, Gaudin RA, Tröltzsch D, Hamm B, Makowski MR, Schüle CY,...
Read more
GALLERY