Development and technical validation of a standardised CT image preprocessing workflow for reproducible AI for forensic anthropological applications.
Authors
Affiliations (3)
Affiliations (3)
- National Forensic Sciences University, Sector 9, Gandhinagar, 382007, Gujarat, India.
- Department of ICT, Pandit Deendayal Energy University, Gandhinagar, India.
- National Forensic Sciences University, Sector 9, Gandhinagar, 382007, Gujarat, India. [email protected].
Abstract
The use of computed tomography (CT scans) images is increasing in forensic anthropology to address ethical, logistical, and preservation challenges related to physical cadaveric collections. However, preparing large volumes of medical images for artificial intelligence (AI) applications remains challenging, particularly in terms of standardisation and data privacy. The aim of this study is to develop an automated preprocessing workflow for CT images to support machine learning (ML) and deep learning (DL) applications in forensic anthropology research and education. A Python-based pipeline was designed in three stages: data de-identification and sorting, patient table removal, and bone windowing. A script was developed for de-identification, and a custom U-Net architecture was developed for both patient table removal and bone windowing. The workflow was developed and applied to CT thorax. Model performance was evaluated using standard metrics and compared with ground-truth models. DICOM metadata de-identification was performed as part of preprocessing; all predefined identifying attributes were successfully processed in the 200 DICOM files examined. The patient table removal model removed the table successfully in 96% of datasets (192/200) and achieved 99.94% accuracy on the original, non-augmented test images. The bone segmentation model achieved a Dice score of 93.12% and a precision of approximately 90%. The segmented 3D models showed high agreement with the ground-truth model, with minimal geometric deviation. We have proposed an integrated workflow that provides an ethical, efficient and standardised solution for de-identifying and preprocessing CT imaging data. It also enables the preprocessing of a scalable dataset for AI applications, thereby supporting the development of methodologies for the modern population, with future expansion to full-body scans and other imaging modalities.