Topogram-based anatomical labelling of CT series: anatomy-aware CT data processing using deep learning.
Authors
Affiliations (9)
Affiliations (9)
- Data Integration Center, Central IT Department, University Hospital Essen, Essen, Germany.
- Institute for Artificial Intelligence in Medicine (IKIM), University Hospital Essen, Essen, Germany.
- Institute of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Essen, Germany.
- Institute for Transfusion Medicine, University Hospital Essen, Essen, Germany.
- Center of Sleep and Telemedicine, University Hospital Essen - Ruhrlandklinik, Essen, Germany.
- Department of Anesthesiology and Operative Intensive Care Medicine (CVK, CCM), Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany.
- Department of Diagnostic Imaging, Oncological Radiotherapy and Hematology, Fondazione Policlinico Universitario "A. Gemelli" IRCCS, Rome, Italy.
- Institute for Artificial Intelligence in Medicine (IKIM), University Hospital Essen, Essen, Germany. [email protected].
- Institute of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Essen, Germany. [email protected].
Abstract
This study provides a Rapid Analysis and Processing of Image Data (RAPID) framework that combines deep learning-based CT topogram analysis with DICOM spatial geometry to enable reliable anatomical labelling of CT series independent of inconsistent textual metadata. In this single-centre retrospective study, three YOLOv8-based models comprising the RAPID framework were trained on CT topograms to perform global anatomical classification, body-region detection, and landmark detection. Classification used 83207 topograms (20,802 test), while landmark and body region detection models were trained on 2000 (500 test) and 1926 (481 test) topograms, respectively, collected between 2003 and 2022. Model performance was evaluated using the F1 score and mAP50, with additional external validation on the external cohort. Furthermore, three radiologists independently reviewed 150 randomly selected predictions for detection models using a Likert-scale-based clinical assessment with inter-rater agreement. Across a total of 65,250 patients (median age, 62 years; interquartile range, 23; 44% female) included in training and testing, inference performance achieved an overall internal F1 score of 0.920 and an external score of 0.970 for classification. Body region and landmark detection achieved internal mAP50 values of 0.993 and 0.958, with corresponding F1 scores of 0.996 and 0.957, respectively. The external mAP scores for detection tasks were 0.952 and 0.926, with corresponding F1 scores of 0.929 and 0.905, respectively. Experts' reviews were generally consistent with the technical evaluation. RAPID enables accurate image-derived anatomical labelling of CT series using topograms. Question Reliable anatomy-based CT series labelling is essential for clinical workflows, but the traditional approach that relies on inconsistent DICOM metadata requires manual review and limits scalability. Findings The three proposed deep learning models achieved high performance in identifying anatomical regions and landmarks, with expert assessments in agreement with the quantitative evaluation. Clinical relevance Deep Learning-based analysis of CT topograms with DICOM-derived spatial geometry, enables reliable and reproducible image-based anatomical labelling of CT series, reducing reliance on inconsistent DICOM attributes and improving data consistency and scalability for clinical applications.