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

P. Bilha Githinji, Xi Yuan, Zhenglin Chen, Ijaz Gul, Dingqi Shang, Wen Liang, Jianming Deng, Dan Zeng, Dongmei Yu, Chenggang Yan, Peiwu QinarXiv 2024 · DOI · arXiv

Abstract

Unsupervised anomaly detection in medical imaging often struggles with the "separability gap" between healthy and pathological samples. We present PopuSense, a novel module that leverages population-level context via hypergraph modeling to refine latent representations. By analyzing intra-group variations within a mini-batch, PopuSense improves pathology detection in contrast-based images like brain MRIs, though challenges remain for texture-heavy modalities.

Introduction

Pathology detection typically relies on unsupervised convolutional models, such as autoencoders, to learn a representation of "healthy" tissue. Anomalies are then identified by high reconstruction errors. However, these models face two primary challenges: the separability gap, where healthy and pathological distributions overlap, and intensity bias, where models perform well on high-contrast images (e.g., MRI) but fail on texture-driven variations (e.g., retinal fundus images).

Standard models focus on local or global spatial context within a single image. This research argues that such approaches miss "intra-group variations"—the subtle differences across a population of healthy subjects—which are essential for building a robust definition of normality.

Methodology

The proposed framework consists of two primary components: PDCCore and the PopuSense Module.

PDCCore

This baseline convolutional autoencoder utilizes a ResNet101 backbone to compress images into a latent code $z$, which is subsequently reconstructed. Training is guided by a combined loss function of Mean Squared Error (MSE) and Structural Similarity Index (SSIM) to ensure both pixel-level accuracy and structural integrity.

PopuSense Module

The core innovation, PopuSense, treats a mini-batch of images as a proxy for a population. It employs a graph-theoretic approach where images are represented as nodes. A hypergraph is constructed to capture higher-order relational associations between similar instances in the batch. A Hypergraph Convolutional Network (HGC) then processes these associations to produce a "population-aware" embedding, which augments and refines the original latent code before reconstruction.

The PopuSense architecture integrates a hypergraph-based refinement module into a convolutional autoencoder framework to capture population-level context.

Experimental Setup and Datasets

The model was evaluated on two distinct medical imaging datasets:

  1. Contrast-based: A public brain tumor MRI dataset (481 healthy images for training).
  2. Texture-based: The IRFundusSet retinal fundus dataset (3,515 healthy images).

Images were standardized to 160x160 resolution and preprocessed using median blur and Contrast Limited Adaptive Histogram Equalization (CLAHE) for noise reduction.

Results and Discussion

Population Context and Separability

Experiments comparing "Narrow" (35% batch neighbors) vs. "Wide" (70% batch neighbors) PopuSense revealed that a wider population context yields superior results. On the brain MRI dataset, the Wide PopuSense configuration achieved the highest performance, showing a statistically significant improvement in the mean residual error difference (8.654) compared to the baseline (4.906). This indicates a significantly improved balance between sensitivity and specificity (AUC).

Reconstruction Fidelity

Qualitative analysis suggests that while vanilla autoencoders tend to "average" features—resulting in blurry outputs—PopuSense effectively "stacks" information from the group. This leads to more detailed reconstructions of healthy anatomy, even when parts of the image are masked.

Qualitative results demonstrating the model's ability to reconstruct structural patterns from the learned population context.

The Texture Challenge

Despite improvements in contrast-based detection, the model struggled to outperform chance on retinal scans. This reinforces the difficulty of texture-based anomaly detection and suggests that current hypergraph modules may require further adaptation to perceive complex textural patterns effectively.

Conclusion

PopuSense demonstrates that incorporating population-level context can significantly enhance the detection of pathologies in high-contrast medical imaging. Future work will focus on adapting the module for texture-based inputs and addressing the challenge of "over-generalization," where models become so proficient they begin to reconstruct the anomalies themselves.

Summary updated

BibTeX
@misc{githinji2024harnessing,
  author = {P. Bilha Githinji and Xi Yuan and Zhenglin Chen and Ijaz Gul and Dingqi Shang and Wen Liang and Jianming Deng and Dan Zeng and Dongmei Yu and Chenggang Yan and Peiwu Qin},
  title = {{Harnessing Intra-group Variations Via a Population-Level Context for Pathology Detection}},
  year = {2024},
  eprint = {2403.02307},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.02307},
}
Figure