Skip to content

AI Workflow Improves CT Scan Preparation for Forensic Anthropology, Study Finds

Researchers at NFSU develop an AI workflow for forensic CT scans, achieving 96% table-removal success and 93.12% Dice overlap in bone segmentation.

By 3 min read

Researchers in India have developed and evaluated an automated workflow for preparing computed tomography (CT) images for forensic anthropology, reporting promising results in removing scanning equipment from images and identifying skeletal structures.

The study, published on November 17, 2026, in the International Journal of Legal Medicine, involved researchers from National Forensic Sciences University in Gandhinagar and Pandit Deendayal Energy University.

The research was conducted by Gargi Jani, Smit Patel, Mohendra Roy, Anjali Sehrawat and Bhoomika M. Patel.

CT scans provide detailed internal images that can be used to examine skeletal anatomy and create three-dimensional representations of bones.

In forensic anthropology, these digital representations can support research into skeletal characteristics while reducing some of the practical and ethical challenges associated with physical human-remains collections.

However, preparing medical imaging data for artificial intelligence requires considerable processing.

CT datasets can contain identifying patient information, imaging equipment and structures that are irrelevant to a particular research objective.

Differences in imaging formats and preparation methods can also complicate the development of consistent training datasets.

The researchers addressed these challenges through an integrated Python-based processing workflow.

The system contains three main stages: de-identifying and organizing imaging files, removing the patient support table and processing images to isolate skeletal structures.

For the automated image-processing stages, the researchers developed a custom U-Net-based deep-learning architecture.

The workflow was evaluated using thoracic CT images.

According to the reported results, all predefined identifying metadata attributes were successfully processed in the 200 DICOM files examined during the de-identification assessment.

The table-removal model successfully removed the scanning table from 192 of 200 datasets, representing a 96% success rate.

The researchers also reported 99.94% accuracy on original, non-augmented test images for the table-removal model.

For bone segmentation, the model achieved a Dice similarity score of 93.12% and precision of approximately 90%.

The Dice score measures the overlap between a predicted segmented region and its reference representation. Higher values generally indicate stronger agreement.

Three-dimensional skeletal models generated through the workflow also showed close agreement with reference models, with limited reported geometric differences.

These findings suggest that automated preprocessing could reduce repetitive preparation work and support the creation of standardized imaging datasets for forensic anthropology research.

However, the reported performance measures describe image processing, not the accuracy of forensic identification.

The study does not establish that the system can independently identify unknown remains, estimate biological characteristics reliably across populations or replace expert anthropological examination.

Its evaluation also focused on thoracic CT images. Additional testing would be necessary to establish performance across other anatomical regions, imaging protocols and datasets.

The authors propose future expansion to full-body scans and other imaging modalities.

The research demonstrates how artificial intelligence may contribute to forensic anthropology by improving the preparation and consistency of digital skeletal data.

Its potential value lies in supporting subsequent scientific research and model development, rather than making independent forensic conclusions.