Harnessing Intra-group Variations Via a Population-Level Context for Pathology Detection

TL;DR

The paper tackles the problem of limited separability between healthy and pathological distributions in pathology-detection autoencoders, especially for texture-based medical images. It introduces PopuSense, a latent-code refinement module that uses a hypergraph to encode population-level context from mini-batches and augment the PDCCore latent space. Across brain-tumor MRI (contrast-based) and retinal fundus imaging (texture-based) datasets, PopuSense improves separability for contrast inputs, particularly with wide neighborhood settings, but shows limited gains on texture-based data. This work suggests a promising direction for incorporating intra-group heterogeneity into pathological detection, while highlighting modality-dependent effectiveness and the need for further tuning across imaging modalities.

Abstract

Realizing sufficient separability between the distributions of healthy and pathological samples is a critical obstacle for pathology detection convolutional models. Moreover, these models exhibit a bias for contrast-based images, with diminished performance on texture-based medical images. This study introduces the notion of a population-level context for pathology detection and employs a graph theoretic approach to model and incorporate it into the latent code of an autoencoder via a refinement module we term PopuSense. PopuSense seeks to capture additional intra-group variations inherent in biomedical data that a local or global context of the convolutional model might miss or smooth out. Proof-of-concept experiments on contrast-based and texture-based images, with minimal adaptation, encounter the existing preference for intensity-based input. Nevertheless, PopuSense demonstrates improved separability in contrast-based images, presenting an additional avenue for refining representations learned by a model.

Figures (7)

• Figure 1: Overview of the population-sensitive latent code refinement approach.
• Figure 2: The latent code of a pathology detection autoecoder is augmented with a population-level context to leverage intricate or hidden associations within a group of similar observations. PopuSense, the corresponding module, takes as input the latent code of the autoencoder and estimates a hypergraph to capture higher-order relationships within a group. Afterward, a hypergraph convolutional neural network learns the associated embedding, which is then incorporated into the original latent code via a single fully connected layer.
• Figure 3: Prototype output. Figure a) is the original input, b) is the reconstruction by the vanilla autoencoder, and c) is the output when PopuSense is utilized. Additional detailing is observed in the outputs in c). Note that the reconstructions are for the normative distribution; anomalous regions should not be reconstructed.
• Figure 4: PDCCore capacity. Column a) corresponds to the original input, b) the augmented input, and c) the reconstructed output. The orange boxes highlight areas with masked augmentation, which are successfully reconstructed in the output images in c).
• Figure 5: Empirical cumulative distribution plots for the error scores, showing the extent of overlap between healthy and pathological distributions for contrast and textural input, and across the three model configurations. Along the x-axis are the corresponding rug plots.