Adversarial Learning Under Hybrid Perturbations for Robust Acute Lymphoblastic Leukemia Classification

Jie Chen, Xinyuan Liu, Xintong Liu, Jianqiang LiAAAI 2025 · DOI

Introduction

Deep learning has revolutionized the classification of Acute Lymphoblastic Leukemia (ALL) cells, offering a path toward faster and more accurate early diagnosis. However, these models often lack the robustness required for clinical deployment. In real-world medical imaging, data is rarely "clean." It is frequently subject to various forms of interference that can lead to catastrophic diagnostic failures.

The Challenge: Brittle Models in Clinical Settings

Standard AI models for leukemia diagnosis are notoriously sensitive to two primary types of interference:

  1. Pixel-level perturbations: Digital noise or graininess inherent in imaging hardware.
  2. Spatial-transformed perturbations: Variations in image capture, such as slight rotations, shifts, or scaling.

Even minor variations—imperceptible to the human eye—can cause a model to flip its prediction from "benign" to "malignant." This brittleness poses a significant risk in medical environments where imaging conditions are inconsistent.

Vulnerability of standard models to pixel-level noise, leading to misclassification.

Methodology: The HPAT-ALLCC Framework

To address these vulnerabilities, researchers developed the Hybrid-Perturbation Adversarial Training (HPAT) framework. This approach employs a "min-max" optimization strategy: the system identifies the most challenging adversarial versions of an image (maximizing loss) and then trains the model to classify them correctly (minimizing loss).

The complete HPAT-ALLCC framework, showing the generation of hybrid adversarial examples and the training process with Mixed Batch Normalization (MixBN).

Technical Innovations

  • STBO (Spatial Transform based on Bayesian Optimization): Traditional methods for finding worst-case spatial transforms are computationally expensive. HPAT uses Bayesian Optimization to efficiently pinpoint the exact rotation or shift that most confuses the model in just a few iterations.
  • Hybrid Adversarial Generation: The framework combines STBO with Projected Gradient Descent (PGD) to create "hybrid" samples that are simultaneously noisy and spatially transformed, reflecting complex real-world conditions.
  • MixBN (Mixed Batch Normalization): Adversarial training often degrades performance on "clean" images. HPAT introduces a dual-branch normalization module (MixBN) that processes clean and adversarial images separately, maintaining high accuracy across both data types.

Experimental Setup

The study utilized the C-NMC dataset from the IEEE ISBI 2019 B-ALL challenge, comprising 12,527 labeled images of benign and malignant cells. The data was split into training (8,258), validation (2,132), and test (1,867) sets. Preprocessing included center-cropping to 300x300 pixels and stain normalization to ensure consistency across the dataset.

Key Results and Insights

The researchers identified EfficientNet-B3 as the optimal backbone architecture, starting with a baseline accuracy of 84.5%.

  • Enhanced Robustness: In environments with both noise and spatial shifts, standard models saw accuracy drop to approximately 48%. The HPAT-trained model demonstrated significantly higher stability.
  • Rate of Invariance (ROI): The HPAT model achieved a superior ROI, meaning its predictions remained consistent despite perturbations, compared to models trained on single-type adversarial defenses.
  • MixBN Effectiveness: The inclusion of Mixed Batch Normalization improved the F1-score from 79.8% to 82.5%, successfully mitigating the trade-off between robustness and clean-data accuracy.

Conclusion

The HPAT-ALLCC framework represents a significant step toward reliable AI-assisted leukemia diagnosis. By defending against hybrid perturbations and utilizing efficient optimization techniques like STBO, the system ensures that diagnostic accuracy remains high even under imperfect imaging conditions. Future work will focus on expanding this robustness to other clinical artifacts like blurring and varying lighting conditions across different medical facilities.

Summary updated

BibTeX
@inproceedings{chen2025adversarial,
  author = {Jie Chen and Xinyuan Liu and Xintong Liu and Jianqiang Li},
  title = {{Adversarial Learning Under Hybrid Perturbations for Robust Acute Lymphoblastic Leukemia Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i2.32204},
}
Figure