A Parameterized Generative Adversarial Network Using Cyclic Projection for Explainable Medical Image Classification

Xiangyu Xiong, Yue Sun, Xiaohong Liu, Chan-Tong Lam, Tong Tong, Hao Chen, Qinquan Gao, Wei Ke, Tao Tan

Introduction

Deep learning's success in medical imaging is often hindered by the scarcity and sensitivity of high-quality labeled data. Small-scale datasets frequently lead to overfitting, where models memorize training samples instead of learning generalizable pathological features. While data augmentation and Generative Adversarial Networks (GANs) are common solutions, they often lack control over the decision boundary or suffer from mode collapse. This paper introduces ParaGAN (Parameterized Generative Adversarial Network), a framework designed to augment medical datasets by anchoring synthetic images to a mathematically defined decision boundary, thereby improving both classification performance and model interpretability.

The ParaGAN Framework

ParaGAN addresses the limitations of standard GANs by incorporating a "projection distance" parameter into a cyclic GAN architecture. The methodology is structured around several key innovations:

Hyperplane Pre-training

The process begins by training an auxiliary classifier ($C_{aux}$) using a hinge loss. This step establishes an "optimal hyperplane" in the feature space, effectively creating a mathematical fence that divides the two classes (e.g., Benign vs. Malignant).

Parameterized Generation and Cyclic Projection

Unlike traditional GANs that rely on random noise, ParaGAN’s generators ($G_{X2Y}$ and $G_{Y2X}$) are conditioned on the distance to this hyperplane. By calculating the required distance to transition from one class to another, the model generates synthetic samples at specific, informative locations in the latent space.

To maintain structural integrity, the model employs a Cyclic Projection mechanism. This includes a cycle-consistency loss and a specific projection distance loss, ensuring that synthetic images remain at the intended distance from the decision boundary while preserving the original image's core features upon reconstruction. Visualization via t-SNE confirms that these synthetic samples effectively fill gaps in the latent space, clustering along the margins of the decision boundary to "strengthen" the classifier.

Weighted Synthetic Loss

To integrate synthetic samples into the final downstream classifier, the authors utilize a weighted synthetic loss. A hyperparameter ($\alpha$) balances the influence of synthetic versus real data, acknowledging that while synthetic images are useful for boundary definition, they may still contain minor artifacts.

Explainability via Class-Difference Maps (CDMs)

Beyond data augmentation, ParaGAN introduces Class-Difference Maps (CDMs) as a more transparent alternative to traditional explainable AI methods like Grad-CAM heatmaps. By projecting a source image directly onto the decision boundary, the model identifies the minimal changes required to alter a classification.

CDMs visualize the specific features—such as the presence of a posterior acoustic shadow in an ultrasound—that drive a diagnostic decision. This tells a clinician not just where the model is looking, but what specific features would need to change to alter the diagnosis.

Experimental Evaluation

The researchers evaluated ParaGAN on two challenging medical imaging tasks:

  • Breast Ultrasound: A combination of the BUSI and UDIAT datasets, focused on distinguishing benign from malignant lesions.
  • COVID-19 CT: The COVID-CT dataset, used to classify non-COVID-19 vs. COVID-19 lung scans.

Key Results

ParaGAN consistently outperformed baseline models, conventional augmentation, and advanced GANs like ACGAN and CycleGAN:

  • Accuracy Boost: On the Breast Ultrasound dataset, ParaGAN achieved an accuracy of 0.895, significantly higher than the original model's 0.849 and conventional augmentation's 0.877.
  • AUC Improvements: On the COVID-CT dataset, it reached an AUC of 0.883, compared to 0.836 for the baseline.
  • Qualitative Superiority: Visually, ParaGAN generated more realistic medical features, such as "ground glass" opacities in COVID-19 lung scans and posterior acoustic shadows in breast ultrasounds, which are critical for clinical diagnosis.

Limitations and Future Work

While the results are promising, the authors identify several areas for future development. Currently, the framework is limited to binary classification; extending this to multi-class scenarios will require handling more complex hyperplane geometries. Additionally, the performance is sensitive to the synthetic weight ($\alpha$), suggesting a need for more automated balancing methods. Finally, the authors plan to conduct clinical validation to determine the practical utility of CDMs for medical professionals in diagnostic settings.

Summary updated

Abstract

Although current data augmentation methods are successful to alleviate the data insufficiency, conventional augmentation are primarily intra-domain while advanced generative adversarial networks (GANs) generate images remaining uncertain, particularly in small-scale datasets. In this paper, we propose a parameterized GAN (ParaGAN) that effectively controls the changes of synthetic samples among domains and highlights the attention regions for downstream classification. Specifically, ParaGAN incorporates projection distance parameters in cyclic projection and projects the source images to the decision boundary to obtain the class-difference maps. Our experiments show that ParaGAN can consistently outperform the existing augmentation methods with explainable classification on two small-scale medical datasets.
BibTeX
@inproceedings{xiong2024parameterized,
  author = {Xiangyu Xiong and Yue Sun and Xiaohong Liu and Chan-Tong Lam and Tong Tong and Hao Chen and Qinquan Gao and Wei Ke and Tao Tan},
  title = {{A Parameterized Generative Adversarial Network Using Cyclic Projection for Explainable Medical Image Classification}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10448260},
}
Figure