Autoassociative Learning of Structural Representations for Modeling and Classification in Medical Imaging

TL;DR

The paper presents Autoassociative Structural Representations (ASR), a neurosymbolic autoencoder that reconstructs images using explicit visual primitives to enforce interpretable object-like explanations. It demonstrates that ASR latent features can improve explainability and downstream classification in medical imaging, outperforming a conventional autoencoder in thyroid histology patch classification. The approach combines a convolutional encoder with a differentiable renderer to enable end-to-end optimization under a pixelwise MMSE loss, and findings highlight the value of coarse scale features for diagnostic tasks. These results support incorporating domain priors through structured primitives to enhance robustness and interpretability in medical image analysis.

Abstract

Deep learning architectures based on convolutional neural networks tend to rely on continuous, smooth features. While this characteristics provides significant robustness and proves useful in many real-world tasks, it is strikingly incompatible with the physical characteristic of the world, which, at the scale in which humans operate, comprises crisp objects, typically representing well-defined categories. This study proposes a class of neurosymbolic systems that learn by reconstructing images in terms of visual primitives and are thus forced to form high-level, structural explanations of them. When applied to the task of diagnosing abnormalities in histological imaging, the method proved superior to a conventional deep learning architecture in terms of classification accuracy, while being more transparent.

Figures (9)

• Figure 1: The overview of ASR's architecture.
• Figure 2: Examples of patches extracted from three different WSIs of the thyroid gland. In the Benign class (a) the follicles are regularly distributed across the tissue and have a circular shape. Aside from the vesicles, not much connective tissue is visible. The Hashimoto's disease (b) features lymphoplasmacytic infiltration, which manifests as numerous dark purple-stained cells (lymphocytes) in the interfollicular area, while the follicles are smaller and sparsely distributed. In Nodularity (c), we observe a small number of thyroid follicles and a high proportion of connective tissue.
• Figure 3: Reconstructions of sample test images produced by the Regularized_2 model.
• Figure 4: The final pruned decision tree induced from the attributes provided by the Base_2 model.
• Figure 5: Example Thyroid Gland Whole Slide Image from the Biospecimen Research Database (BRD) BiospecimenResearchDatabase.