X-ray hyperspectral segmentation taps persistent homology: H1 loops on UMAP superpixel clouds lift pixel F1 from 0.8245 to 0.8945
Synopsis
Segmentation of multi-energy X-ray transmission hyperspectral images has relied on convolutional networks for local spectral-spatial features, without explicitly using the geometry of the embedded manifold. This work computes Vietoris-Rips persistent homology on each superpixel's spectral point cloud after UMAP projection, extracts H1 loop statistics, processes them with a graph convolutional network, and broadcasts the result back to every pixel for concatenation with pixel-level CNN features. On the cigarette dataset, the broadcasting hybrid reaches pixel F1 0.8945 and superpixel F1 0.9329, above the CNN-only pixel F1 of 0.8245.
Interpretation
Segmentation of multi-energy X-ray transmission hyperspectral images can now explicitly use the topology of the spectral manifold rather than relying only on local spectral-spatial convolutional features. The prior CNN baselines captured only local patterns in the UMAP embedding and did not characterize the global structure of the spectral manifold; the pixel-wise CNN reached pixel F1 0.8245 and the superpixel-wise CNN 0.8144. On the cigarette dataset, the broadcasting hybrid reaches pixel F1 0.8945 and superpixel F1 0.9329, 0.070 and 0.086 above the best CNN baseline; the pooling hybrid also beats both CNN variants but trails broadcasting by 0.061 in superpixel F1.
The approach turns each superpixel's spectral point cloud into a set of topological statistics, then lets a graph network send that region-level information back to pixels. Persistent homology needs enough points to form stable topological structures, and computing it per pixel is impractical, so the pipeline first runs SLIC superpixel segmentation on the UMAP projection and then computes Vietoris-Rips persistent homology on each superpixel's point cloud. Each superpixel node starts as a 60-dimensional vector of 30 candidate topological descriptors, 25 spectral statistics and 5 geometric descriptors; after ablation the graph network keeps only 11 H1 statistics, which with 30 spatial descriptors form 41-dimensional node features fed to a three-layer graph convolutional network.
The topological gain appears in both the original 16-band spectral space and the 5-dimensional UMAP space, indicating the cyclic structures are not created by the embedding. In the graph-only ablation, adding the 11 H1 features raises target-class F1 from 0.685 to 0.764 on raw hyperspectral data and from 0.778 to 0.823 in UMAP space. The same ablation shows that expanding from 11 to 19 or 30 topological features lowers raw-hyperspectral F1 to 0.752 and 0.725; with 30 topological features alone F1 is only 0.283, because persistent homology is invariant to pixel permutations within a superpixel and discards absolute spectral intensities and spatial coordinates.
Perspective
The demonstrated scope is limited to multi-energy X-ray transmission image patches from the cigarette dataset, a binary segmentation task separating cigarette from luggage background, evaluated by pixel-level and superpixel-level accuracy and F1.
Persistent homology is invariant to pixel permutations within a superpixel, so used alone it discards absolute spectral intensities and spatial coordinates, and the 30 topological features reach only F1 0.283; the 11-H1 subset was selected in an ablation with the CNN branch disabled, and expanding to 19 or 30 features lowers F1.
