Leveraging INRs’ Initialization for Faster 3D Unsupervised Medical Image Anomaly Detection

Common approaches to unsupervised medical anomaly detection rely on generative models that learn the distribution of normal images and associate high reconstruction errors with anomalies. One limitation of such solutions is that the inherent models process images as discrete voxel grids, resulting in a significant complexity and memory footprint, particularly for 3D volumes. Implicit Neural Representations offer an alternative by modeling images as continuous implicit functions, with small MLPs that process voxels independently, especially reducing memory requirements. However, only a few approaches have explored INRs for anomaly detection, and existing methods based on autodecoders have overlooked the test-time optimization process. This optimization often requires numerous iterations to reach a suitable latent vector for each test sample, resulting in a lengthy inference process. In this work, we propose and evaluate different strategies to initialize the test latent vectors, effectively accelerating the optimization convergence and reducing inference time. Experimental results in the context of 3D brain anomaly detection demonstrate statistically significant time reductions compared to state-of-the-art methods, while achieving equal or superior performance in terms of anomaly segmentation performance.