AFV fibre model V2
A 3D model for detecting papyrus fibres in CT volumes and estimating their local orientation. It distinguishes horizontal (H) and vertical (V) fibre families.
The model uses a residual-encoder nnU-Net with a single grayscale CT input and additional heads for fibre centres, local affinities and direction. It continues the AFV ~9 µm fibre model.
Scan resolution
Optimized for CT sampled at approximately 9 µm per voxel. Training inputs use 8.64–9.6 µm voxel spacing, including downsampled levels of finer-resolution scans. Prepare input volumes at this scale.
PHerc1447, 8.64 µm per voxel. Two central slices of the same 128³ input; colour intensity shows the predicted probability for each fibre family. Qualitative example from a training scroll.
Outputs
- Fibre segmentation: four voxelwise logits, ordered background, vertical, horizontal, intersection. Apply softmax across the four channels to obtain probabilities. The intersection channel is retained from the original architecture; its accuracy has not been established separately.
- Direction: three components in XYZ describing the local fibre axis. Normalize the vector before use; opposite signs represent the same orientation.
- Auxiliary heads: one fibre-centre channel and six local-affinity channels are also included in the checkpoint.
Compression augmentation
Training includes synthetic geometric compression to expose the model to tightly packed papyrus layers. A smooth deformation along the local layer normal first narrows the spaces between layers; stronger variants also compress the material itself. The CT and fibre annotations follow the same deformation, keeping the targets aligned with the image.
Recorded augmentation examples from the same PHerc0175A region, with a shared centre and display contrast. Compression here simulates physical compaction of the papyrus.
Training data
Training used regions from these 30 scrolls:
PHerc0125, PHerc0139, PHerc0172, PHerc0175A, PHerc0175B, PHerc0191, PHerc0211, PHerc0257, PHerc0268, PHerc0306B, PHerc0332, PHerc0343, PHerc0343P, PHerc0358, PHerc0483A, PHerc0483B, PHerc0490A, PHerc0490B, PHerc0500P2, PHerc0800, PHerc0813, PHerc0814, PHerc0826, PHerc0841, PHerc0846A, PHerc0846B, PHerc1203, PHerc1218, PHerc1447 and PHercParis4.