Anomaly-Aware Multiple Instance Learning for Rare Anemia Disorder Classification

Salome Kazeminia, Ario Sadafi, Asya Makhro, Anna Bogdanova, Shadi Albarqouni, Carsten Marr

Abstract

Diagnosing rare hereditary hemolytic anemias (HHAs) from blood samples is a challenging "needle-in-a-haystack" task due to the scarcity of hallmark morphological deformities. This paper introduces an anomaly-aware Multiple Instance Learning (MIL) framework that models the distribution of healthy red blood cells to identify anomalous cells. By fusing statistical anomaly scores with learned attention weights, the proposed model achieves higher accuracy and superior explainability compared to standard attention-based methods, particularly in data-scarce clinical scenarios.

Introduction

Diagnosing rare hereditary hemolytic anemias (HHAs)—such as sickle cell disease (SCD) or thalassemia—requires identifying a tiny fraction of deformed red blood cells among thousands of healthy ones. Traditional supervised deep learning is often impractical for this task because it requires expensive, error-prone cell-level labeling. While Multiple Instance Learning (MIL) allows for "bag-level" labels (e.g., patient-level diagnosis), standard MIL often fails when the disease signal is weak or training data is scarce. The few diseased cells are frequently "washed out" by the majority of healthy cells, and complex attention mechanisms struggle to converge with limited samples.

Methodology

To address these limitations, the authors propose an Anomaly-aware Pooling mechanism that incorporates an explicit measure of how much a cell deviates from a healthy norm.

The proposed anomaly-aware MIL architecture, featuring feature extraction, GMM-based healthy distribution modeling, and hybrid attention-anomaly pooling.

Pipeline Components

  1. Feature Extraction: A Mask R-CNN detects individual cells, which are then processed by a 4-layer CNN encoder to generate latent representations.
  2. Healthy Distribution Modeling: A Gaussian Mixture Model (GMM) is fitted to the features of cells from healthy individuals (negative bags) to define the "normal" cluster.
  3. Anomaly Scoring: For each new cell, the model calculates the Mahalanobis distance from the healthy GMM cluster, providing a statistical measure of its "weirdness."
  4. Hybrid Pooling: The final importance weight for a cell is a fusion of a standard learned Attention score and the statistical Anomaly score, combined via a $1 \times 1$ convolution layer.
  5. Optimization: A "Single Instance Classifier" (SIC) loss is used in early training epochs to jumpstart the model before transitioning to the full MIL loss.

Experimental Setup

The study utilized 3,630 microscopic images from 71 individuals in the CoMMiTMenT study. The dataset included four disease classes—Sickle Cell Disease (SCD), Thalassemia (Thal), Hereditary Xerocytosis (Xero), and Hereditary Spherocytosis (HS)—alongside a healthy control group.

Results and Discussion

The proposed "Anomaly + Att + SIC" method outperformed standard attention-based MIL across key metrics, achieving an accuracy of 0.787 (vs. 0.752) and an AU RoC of 0.854 (vs. 0.826).

Visual comparison of Anomaly vs. Attention scores across different disease types. The anomaly mechanism correctly flags irregular cells that standard attention often misses.

Key Insights

  • Leveraging the "Normal": By explicitly modeling what "healthy" looks like, the model can identify a wide variety of unhealthy deviations, even with limited disease-specific samples.
  • Explainability: Visualization of the anomaly scores provides a clear heatmap for clinicians. In SCD samples, the anomaly mechanism correctly flagged irregular, elongated cells that standard attention mechanisms often overlooked.
  • Fusion is Essential: Ablation studies showed that using only anomaly scores performed poorly (0.630 accuracy), confirming that the synergy between learned attention and statistical distance is the primary driver of performance.

Conclusion

The anomaly-aware MIL framework provides a robust solution for rare disease classification where data is limited and hallmark features are sparse. While the current model uses a simple GMM, future work could incorporate more advanced out-of-distribution (OOD) detection methods and uncertainty estimation to further enhance clinical utility. Testing on larger, multi-center datasets remains a necessary step for real-world deployment.

Summary updated

BibTeX
@inproceedings{kazeminia2022anomaly,
  author = {Salome Kazeminia and Ario Sadafi and Asya Makhro and Anna Bogdanova and Shadi Albarqouni and Carsten Marr},
  title = {{Anomaly-Aware Multiple Instance Learning for Rare Anemia Disorder Classification}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  doi = {10.1007/978-3-031-16452-1\_33},
}
Figure