Radiographic analysis and virtual removal of speleothemic deposits
The specimen was brought to the University Hospital Centre “Zagreb” (Zagreb, Croatia) and imaged with a clinical computed tomography (CT) scanner (Light Speed Ultra, General Electric, Chicago, USA) using 8 × 1.25 mm collimation, 0.75 pitch, and 120 kVp. Image contrast was considered insufficient for segmentation in the preview on the workstation, and no reconstructions were performed. A different clinical CT scanner (Sensation 40, Siemens Healthineers, Erlangen, Germany) with 40 × 0.6 mm collimation, 0.9 pitch, and 140 kVp at 390 mAs was used for a second scan, producing substantially better contrast (Fig. 3a). Axial sections were reconstructed with a bone weighted kernel (B70s), a slice thickness of 0.6 mm, and an increment of 0.3 mm. A total of 682 slices was generated for coverage of the entire sample.
Nevertheless, segmentation of the incrusted bone remained a particular challenge due to the sometimes very similar densities of the partially fossilised bone fragment and surrounding concretions (Fig. 3a). Hounsfield units (HU) threshold-based or region-growing-based segmentation algorithms were not practicable. A combined, manual and HU threshold-based segmentation protocol, newly developed for this project by one of the authors (PE), was therefore applied. For later validation, segmentation was independently carried out by three readers (R1–3) using the dedicated open-source medical DICOM image viewing software (Horos, v3.2.1, www.horosproject.org): First, in an axially reconstructed series of 702 slices, a free-form region of interest (ROI), most closely corresponding to the demarcation of the bone fragment, was manually drawn on every tenth slice. Particular care was taken not to include components of the calcite crust or omit parts of the bone fragment; furthermore, ROIs were drawn clockwise starting at a twelve o’clock position (Fig. 3b). Second, missing ROIs were interpolated on the remaining slices. After this step, added ROIs were checked by scrolling through the entire data set and manually corrected where necessary. Once each slice contained a ROI circumscribing the bone fragment’s delimitation, the “ROI Volume/Compute Volume...” function allowed a first visual inspection of the virtually cleaned bone fragment. Third, the contents of every ROI on each slice, i.e., the actual bone fragment, was deleted with the “ROI Volume/Erase Content” function, and the resulting data set was saved as a new DICOM series (Fig. 3c). The newly generated and stored template DICOM series, only representing the calcite concrement, was then subtracted from the original (unaltered) axially reconstructed series, resulting in a dataset only showing containing the virtually cleaned femoral fragment (Fig. 3d). Last, using conventional HU threshold segmentation, a three-dimensional (3D) surface rendering was created and exported in the “standard tessellation language” file format for subsequent evaluation and dimensioning with the 3D design software (Rhinoceros 3D, Version 5.4.1, Robert McNeel & Associates, Seattle, WA, USA) (Fig. 3e).
Taxonomic assessment
The taxonomic assessment was performed by a zooarchaeologist (SR), based on measurements and renderings of the virtual 3D models resulting from CT image segmentation. Measures were taken following von den Driesch [4], and results were then compared with published metric data for species in question. Anatomical terminology used follows standards referenced in the Nomina Anatomica Veterinaria (5th edition) [5].
14C radiocarbon dating
A direct radiocarbon date of a small sample of the exposed shaft was conducted in conventional 14C age, and fraction corrected using AMS 13C. Calibrated age ranges using Intcal13.14c (1 sigma) [6] were 3637–3627 (probability distribution 0.133), 3591–3527 (probability distribution 0.867), (2 sigma) 3647–3516 (probability distribution 0.976), 3408–3405 (probability distribution 0.003) and 3398–3384 (probability distribution 0.020).
Statistical analysis
The resulting 3D models of R1–3 were later also statistically compared, by point-to-point comparison of the respective point-clouds, using the dedicated open-source software (CloudCompare V2.6, www.cloudcompare.org) to validate the accuracy of the resulting models (Fig. 4a). For statistical comparison of the inter-reader correlation, the number of points of the point-clouds was reduced to 1 million points per model. Descriptive statistical analysis and boxplots were created using the dedicated software (SPSS Statistics, Version 24, IBM, Armonk, NY, USA), after importing comparative statistical from CloudCompare in the “comma-separated values” data format.