Topology-Preserving Neural Architecture Search for Dense Prediction in Underexplored Modalities

Combining the NAS survey by Elsken et al. (2022) with Fu & Nelson’s (2021) topological regularization, this idea proposes a NAS framework where candidate architectures are evaluated not only for accuracy and efficiency but also for their ability to preserve topological properties—such as connectivity and component count—of the segmented output. This is especially valuable for novel or underexplored modalities (like event-based or thermal imaging) where ground-truth annotations are sparse and topological structure is critical (e.g., medical blood vessel segmentation, as in Sandeep et al., 2025). By explicitly encoding topology into the architecture search objective, we can discover models that are not only performant in traditional metrics, but also produce outputs with meaningful and reliable structure, paving the way for dense prediction in new sensor domains.

References:

1. Enhanced retinal blood vessel segmentation via loss balancing in dense generative adversarial networks with quick attention mechanisms.. D. Sandeep, K. Baranitharan, A. Padmavathi, Loganathan Guganathan (2025). International ophtalmology.
2. Neural Architecture Search for Dense Prediction Tasks in Computer Vision. T. Elsken, Arber Zela, J. H. Metzen, B. Staffler, T. Brox, Abhinav Valada, F. Hutter (2022). International Journal of Computer Vision.
3. Topological Regularization for Dense Prediction. Deqing Fu, Bradley J. Nelson (2021). International Conference on Machine Learning and Applications.
4. Neural Architecture Search for Dense Prediction Tasks in Computer Vision. T. Elsken, Arber Zela, J. H. Metzen, B. Staffler, T. Brox, Abhinav Valada, F. Hutter (2022). International Journal of Computer Vision.