Mainz 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
Q: Fachverband Quantenoptik und Photonik
Q 74: Quantum Information – Concepts and Methods
Q 74.8: Talk
Friday, March 6, 2026, 12:45–13:00, P 10
Exploring Multi-class Image Segmentation Through Localization Phenomena — •Akshaya Srinivasan1,2, Yannick Werner2,3, Alexander Geng1, Berta Garcia Heras3, Ali Moghiseh1, Alexey Bochkarev2, and Maximilian Kiefer-Emmanouilidis2,3 — 1Fraunhofer ITWM, Kaiserslautern — 2RPTU Kaiserslautern-Landau, Kaisersluatern — 3DFKI, Kaiserslautern
We propose an unsupervised, quantum-inspired method for multi-class segmentation of abdominal CT scans based on Anderson localization in image-derived 2D lattice Hamiltonians. Each CT slice is mapped onto a lattice in which pixel intensities define a disordered on-site potential, while nearest-neighbor hopping terms are set by a Gaussian similarity kernel that encodes local image structure. This formulation induces Anderson-like localization of eigenstates driven by contrast variations across the image. Diagonalization of the resulting single-particle Hamiltonian enables segmentation by binning eigenmodes according to their localization lengths, which naturally correspond to anatomical scales. Distinct clusters of localized states align with major anatomical regions, including liver, pancreas, kidneys, and background, producing coherent multi-label segmentation masks without annotated data, pre-processing, or model training. Validation on clinical abdominal CT datasets demonstrates robust performance under varying contrast and noise conditions. The framework is purely linear-algebraic and highlights the potential of Hamiltonian-based models and localization physics for interpretable, physics-driven medical image analysis and quantum-inspired computer vision algorithms.
Keywords: Anderson localization; Quantum-inspired algorithms; Unsupervised learning; Image segmentation; Disordered systems