Mesopotamian Journal of Artificial Intelligence in Healthcare
Abstract
Three-dimensional (3D) surface meshes of anatomical structures are increasingly used for clinical decision support and forensic identification. However, their irregular connectivity, the absence of a canonical vertex ordering, and the scarcity of annotated medical data make them difficult to analyse with deep learning. We present MeshTraveler, a framework that serializes a mesh into a deterministic, geodesically coherent vertex sequence and classifies it with a compact attention-based network. Vertex ordering is formulated as a Traveling Salesman Problem (TSP) over exact geodesic distances and solved with the Nearest Insertion heuristic, so that every vertex is visited exactly once and consecutive vertices remain close along the surface. Each mesh is thus encoded as an L × 3 matrix of vertex coordinates, without requiring manifold meshes, subdivision, or remeshing. The network combines a pointwise residual embedding, four stacked offset-attention layers that model long-range dependencies between vertices, and dual max/average pooling. On the IntrA intracranial aneurysm dataset, MeshTraveler achieves a macro F1-score of 0.9635 using only 500 vertices, compared with 0.9044 for the best reported point-based baseline. On a forensic 3D teeth dataset, it reaches 99.15% accuracy for 16-class tooth identification, and it attains 100% accuracy on the SHREC'11 benchmark, matching the best mesh-specific methods. An ablation study shows that the geodesic TSP ordering improves performance by approximately 2–3 percentage points over unordered vertex input on all datasets, and a comparison of insertion heuristics shows that Nearest Insertion provides the shortest construction time. These results indicate that geodesic TSP serialization is a compact, reproducible, and accurate input representation for medical and forensic mesh classification.
Recommended Citation
Kassimi, Amine; Mouncif, Hamza; Jittou, Asmaa; ElFazazy, Khalid; Gardelle, Thierry Bertin; Mahraz, Mohammed Adnane; Yahyaouy, Ali; Tairi, Hamid; and Riffi, Jamal
(2026)
"MeshTraveler: TSP-Guided Attention for Robust 3D Medical Mesh Analysis,"
Mesopotamian Journal of Artificial Intelligence in Healthcare: Vol. 4:
Iss.
1, Article 11.