Abstract
In this paper, we tackle the task of automatically analyzing 3D volumetric scans obtained from computed tomography (CT) devices. In particular, we address a particular task for which data is very limited: the segmentation of ancient Egyptian mummies CT scans. We aim at digitally unwrapping the mummy and identify different segments such as body, bandages and jewelry. The problem is complex because of the lack of annotated data for the different semantic regions to segment, thus discouraging the use of strongly supervised approaches. We, therefore, propose a weakly supervised and efficient interactive segmentation method to solve this challenging problem. After segmenting the wrapped mummy from its exterior region using histogram analysis and template matching, we first design a voxel distance measure to find an approximate solution for the body and bandage segments. Here, we use geodesic distances since voxel features as well as spatial relationship among voxels is incorporated in this measure. Next, we refine the solution using a GrabCut based segmentation together with a tracking method on the slices of the scan that assigns labels to different regions in the volume, using limited supervision in the form of scribbles drawn by the user. The efficiency of the proposed method is demonstrated using visualizations and validated through quantitative measures and qualitative unwrapping of the mummy.
Introduction
Archaeologists have moved away from physical unwrapping of mummies because the process is destructive and irreversible. While 3D CT scans offer a non-invasive alternative, analyzing them is a significant challenge for computer vision. Unlike modern medical CT scans, mummy data is "noisy": tissues have degraded over millennia, and the radiodensity (Hounsfield Units) of bandages is often nearly identical to the underlying skin. Furthermore, there are no massive, labeled datasets of mummies to train deep learning models. Manual annotation is too time-consuming for experts, and unsupervised methods often fail because the body isn't always the most "salient" part of the scan.
Methodology
The proposed framework is a hierarchical, two-stage approach that minimizes the need for human intervention by leveraging geometric priors.
1. Unsupervised Pre-processing & Geodesic Initialisation
The system first performs exterior removal using histogram analysis to strip away air and template matching (plus Hough transforms) to identify and remove wooden supports. High-intensity voxels representing jewelry or amulets are identified via simple thresholding.
The core innovation lies in the use of geodesic distance. Since bandages and skin look similar in density, the system leverages the spatial structure of the mummy. By calculating the shortest path (geodesic distance) from the exterior into the volume, an initial separation is achieved. Because the body is physically "deeper" than the bandages, the gradient in these distances allows for an approximate separation of the two layers.
2. Weakly Supervised Refinement
The initial segmentation is refined using a 3D version of the GrabCut algorithm. To ensure consistency across the hundreds of 2D slices, a tracking method is employed. If a segment cannot be tracked across successive frames, it is discarded as noise. Finally, users provide "scribbles" only at the final stage to resolve remaining ambiguities, making the process significantly more efficient than full manual labeling.
Datasets and Validation
The primary data source was a 3D CT scan of a human male mummy from the Museo Egizio in Turin, Italy. To validate robustness despite limited data, the researchers used Thin-Plate Spline (TPS) warping to generate synthetic mummy volumes with varying shapes.
Results and Discussion
The proposed method achieved an Intersection over Union (IOU) of 0.81, significantly outperforming standard GrabCut (0.50) and unsupervised saliency-based methods (0.46). The inclusion of the tracking step improved the average IOU by approximately 2-3% and provided smoother results.
Qualitatively, the method successfully produced clear 3D visualizations of the "unwrapped" body, isolating the skeleton and metallic jewelry hidden within the layers. A key insight is that topology is more reliable than radiodensity in this context; the physical fact that bandages must be outside the body is a stronger feature than pixel intensity.
Limitations and Future Work
Despite its success, the method faces challenges with metal artifacts (streaking) from gold jewelry, which can distort nearby tissue density. Additionally, "bleeding" between segments can occur where bandages are pressed tightly against the skin. Future work will involve validating the system on a wider variety of mummification styles across different dynasties and regions.
Summary updated