Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

TL;DR

A parameterized GAN (ParaGAN) that effectively controls the changes of synthetic samples among domains and highlights the attention regions for downstream classification is proposed that can consistently outperform the existing augmentation methods with explainable classification on two small-scale medical datasets.

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.

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

TL;DR

A parameterized GAN (ParaGAN) that effectively controls the changes of synthetic samples among domains and highlights the attention regions for downstream classification is proposed that can consistently outperform the existing augmentation methods with explainable classification on two small-scale medical datasets.

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.
Paper Structure (15 sections, 7 equations, 4 figures, 2 tables)

This paper contains 15 sections, 7 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Dataset Acquisition and Evaluation Metrics
  4. Parameterized Generative Adversarial Network
  5. Objective Function
  6. Downstream Classifier Optimization
  7. Implementation
  8. Network Architecture
  9. Training Settings
  10. Experiments and Results
  11. Experiments on Breast Ultrasound and COVID-CT
  12. Ablation Study
  13. Distribution of Training Samples
  14. Interpretability of Classification
  15. Conclusion and Future Works

Figures (4)

  • Figure 1: Overview of ParaGAN enabling cyclic parameterized projection. (a) Pre-train an auxiliary classifier $C_{aux}$ by hinge loss to provide a hyperplane. (b) The generators translate source images conditioned on the target images' projection distances in forward path, and vice versa for reconstructing sources. $C_{aux}$ reconstructs the projection distances from synthetic images.
  • Figure 2: Qualitative results over mixed Breast Ultrasound and COVID-CT. The synthetic images by the ParaGAN are clearly closer to target domain than that by the CycleGAN cyclegan. The projected results on the hyperplane for source images are displayed.
  • Figure 3: Training samples distributions of various methods for downstream ConvNext visualized by t-SNE van2008visualizing. The proposed ParaGAN tends to fill the latent space and to be along the margins of the hyperplane.
  • Figure 4: The difference between the source images and their projections on a hyperplane can highlight class-specific region, which cannot be deduced from the Grad-CAM gradcam.