Robustly Optimized Deep Feature Decoupling Network for Fatty Liver Diseases Detection

TL;DR

Fatty liver ultrasound classification faces data scarcity and imbalanced class distributions, leading to unreliable detection across stages. The authors propose a two-pronged approach: an Iterative Compression Feature Decoupler (ICFDNet) to separate common and disease-specific features, and Accuracy-based Adaptive Adversarial Training that uses class-aware perturbations $\epsilon_i$ and TRADES-inspired weights $\beta_i$ in a jointly optimized loss $\mathcal{L}$. The objective combines decoupling supervision, a BCELoss term, and a robust adversarial term via $\mathcal{L}=\lambda_1\mathcal{L}_c+\lambda_2\mathcal{L}_s+\lambda_3\mathcal{L}_{at}$ with a Laplacian-based boundary loss $\mathcal{L}_c=\sqrt{(\Delta^2(x)-\Delta^2(\hat{x}))+\xi^2}+\mathcal{L}_1$. On a dataset of 1265 elderly patients, the method achieves an average accuracy of $82.95\%$, a $4.16\%$ gain over the next-best approach, and an $11.67\%$ reduction in best–worst gap, with pronounced improvements on mild/moderate cases, and is presented as a general framework that can boost robustness for other classifiers.

Abstract

Current medical image classification efforts mainly aim for higher average performance, often neglecting the balance between different classes. This can lead to significant differences in recognition accuracy between classes and obvious recognition weaknesses. Without the support of massive data, deep learning faces challenges in fine-grained classification of fatty liver. In this paper, we propose an innovative deep learning framework that combines feature decoupling and adaptive adversarial training. Firstly, we employ two iteratively compressed decouplers to supervised decouple common features and specific features related to fatty liver in abdominal ultrasound images. Subsequently, the decoupled features are concatenated with the original image after transforming the color space and are fed into the classifier. During adversarial training, we adaptively adjust the perturbation and balance the adversarial strength by the accuracy of each class. The model will eliminate recognition weaknesses by correctly classifying adversarial samples, thus improving recognition robustness. Finally, the accuracy of our method improved by 4.16%, achieving 82.95%. As demonstrated by extensive experiments, our method is a generalized learning framework that can be directly used to eliminate the recognition weaknesses of any classifier while improving its average performance. Code is available at https://github.com/HP-ML/MICCAI2024.

Figures (3)

• Figure 1: From left to right, the four ultrasound images are:(1) normal liver, (2) mild, (3) moderate, and (4) severe fatty liver. Squares mark the approximate location of the diaphragm and blood vessels. Arrows point to suspected locations of fat accumulation.
• Figure 2: Overview of the proposed framework. (1) For the feature decoupling phase, we utilize the proposed ICFDNet to generate decoupled features and supervise the decoupling process in different methods. (2) For the adversarial training phase, we use an adaptive adversarial process to add corresponding perturbations to different images, and concatenate them with the decoupled features.
• Figure 3: The network structure of ICFDNet. The network structure of ICFDNet utilizes a U shaped architecture, with each of its ICBlocks, adopting a mini U shaped structure. This design ensures that the network undergoes a complete compression-recovery process during each operation, aiming to refine key features repeatedly.