Abstract
Detection of fatty liver disease via ultrasound is often hindered by image noise and "recognition weakness," where models fail to identify specific disease stages despite high average accuracy. This paper presents ICFDNet, a novel framework utilizing an iterative compression feature decoupling architecture and an accuracy-based adaptive adversarial training strategy. Experimental results on a clinical dataset of 1,265 patients demonstrate an average accuracy of 82.95% and a significant reduction in performance variance across disease stages.
1. Introduction
Non-alcoholic fatty liver disease (NAFLD) is a prevalent condition that is often reversible if detected early. Ultrasound imaging is the primary screening tool; however, it is characterized by low signal-to-noise ratios and subtle textural differences between stages (Normal, Mild, Moderate, and Severe). Standard deep learning models often struggle with these nuances, leading to "recognition weaknesses"—high performance on majority classes but failure on critical intermediate stages like "mild" fatty liver. This study introduces the Iterative Compression Feature Decoupler (ICFDNet) to address these challenges through architectural innovation and robust optimization.
2. Methodology
The proposed framework addresses both feature extraction and model robustness through three primary components.
2.1 Iterative Compression Feature Decoupler (ICFDNet)
Traditional convolutional neural networks often fail to capture the subtle pathological markers in noisy ultrasound images. ICFDNet employs a "U-within-a-U" architecture. Each Iterative Compression Block (ICBlock) functions as a localized encoder-decoder structure, integrating multi-head self-attention and dilated convolutions. By iteratively refining features, the network effectively "decouples" essential pathological information from hardware-induced noise and common anatomical structures.
2.2 YCbCr Color Space Transformation
To further mitigate the impact of noise, the framework transforms input images from the standard RGB color space to YCbCr. This separation of luminance (Y) from chrominance (Cb, Cr) allows the model to prioritize structural and intensity-based information, which is more relevant for ultrasound interpretation than the artificial color mapping often applied by imaging hardware.
2.3 Accuracy-based Adaptive Adversarial Training
The "Robust Optimization" aspect of the network is realized through an adaptive adversarial training strategy. Unlike static adversarial training, this method dynamically adjusts the perturbation level ($\epsilon$) and regularization parameter ($\beta$) based on the model’s real-time accuracy for each specific class.
- For high-performing classes: The model introduces harder adversarial samples to sharpen decision boundaries.
- For low-performing classes: The model prioritizes natural samples to stabilize the learning process. This approach ensures that the model does not achieve high average accuracy at the expense of specific disease stages.
3. Experimental Results
The framework was evaluated using a dataset of 1,265 elderly patients from the Tiaodenghe Community in Chengdu, China. Labels were verified by two senior physicians to ensure ground-truth reliability.
3.1 Comparative Performance
ICFDNet achieved an average accuracy of 82.95%, outperforming established architectures such as ResNet, SENet, and Vision Transformers (ViT). Notably, the model outperformed the second-best performing model (CBAM) by 4.16%.
3.2 Mitigation of Recognition Weakness
A key metric in this study is the "Best-Worst" gap—the difference in accuracy between the best-performing and worst-performing classes. ICFDNet reduced this gap to 18.98%. In contrast, the Vision Transformer (ViT) exhibited a massive 74.12% gap, and ResNet maintained a 32.01% gap. This indicates that ICFDNet provides a much more reliable and balanced diagnostic tool across all stages of fatty liver disease.
4. Discussion and Conclusion
The success of ICFDNet highlights the importance of iterative refinement in medical imaging. By treating adversarial training as a class-wise optimization tool rather than just a security measure, the authors have created a framework that promotes fairness in diagnostic performance. While there is a slight trade-off where balancing classes may marginally affect the accuracy of the "easiest" cases, the overall reliability of the system is significantly improved for clinical application. Future work will focus on refining the optimization process to eliminate this trade-off and extending the framework to other modalities such as CT and MRI.
Summary updated