Speaker
Description
We introduce a new application of variational quantum models in which the structure is designed to reflect domain-specific spatial information in the data. Rather than treating an image as an unstructured collection of features, our approach represents different spatial regions as quantum subsystems and models their interactions through a parametrised Hamiltonian. The main property of our model is that it has a built-in form of interpretability: after training, the learned interaction coefficients can be extracted as region-level relevance scores, indicating which parts of the spatial environment contribute most strongly to the classification.
We apply this approach to image-based cancer mutation classification, focusing on the task of distinguishing KRAS and CTNNB1 mutations from tumour images centred on individual cancerous cells. In numerical experiments, the model achieves strong classification performance and identifies the target cell itself as the most relevant contributor to the prediction. These results suggest that Hamiltonian-structured quantum models can provide a natural way to combine spatial inductive bias with interpretable classification, without requiring additional post-hoc explanation methods or extra evaluation cost.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Hybrid quantum-classical approaches for learning and optimization |