AFSC: Adaptive Fourier Space Compression for Anomaly Detection

Haote Xu, Yunlong Zhang, Liyan Sun, Chenxin Li, Yue Huang, Xinghao DingIEEE Trans. Industrial Informatics 2022 · DOI · OpenReview

TL;DR

This paper introduces AFSC, a novel method for anomaly detection in medical images that leverages adaptive compression of Fourier space information. By focusing on essential healthy image features in the frequency domain, AFSC significantly improves the performance of existing pseudo-health generation-based anomaly detection techniques, leading to more accurate and generalizable lesion detection.

What Problem Does This Paper Address?

Anomaly Detection (AD) in medical imaging is crucial for identifying diseases, especially emergent ones, where labeled data is scarce. Traditional supervised methods struggle with this data deficiency. Pseudo-health generation approaches aim to overcome this by training models to reconstruct healthy images from anomalous inputs. These methods typically work by generating "pseudo-healthy" images and finding anomalies as the difference between the input and the generated output. However, existing methods often fail to capture the true complexity and irregularity of real lesions, limiting their generalization capabilities. This paper addresses the challenge of creating more robust pseudo-healthy images by exploring intrinsic data properties within the frequency domain, specifically by adaptively compressing Fourier space information.

Methods

The core of the proposed method is the Adaptive Fourier Space Compression module (AFSCm). Instead of operating directly on the image space, AFSCm transforms the input image into its frequency domain representation, separating it into amplitude and phase components. It then applies adaptive masks (MA and MP) to compress these components, aiming to retain the most critical information for generating pseudo-healthy images while filtering out lesion-specific noise. This compressed Fourier representation is then used by a generator network to create pseudo-healthy images. The generator is trained using a combination of reconstruction loss, adversarial loss (to make generated images indistinguishable from real healthy images), and a sparsity loss to encourage efficient masks. The paper also details how these masks are learned through an optimization process that approximates a differentiable sigmoid function.

Datasets

The experiments were conducted on three different brain MRI datasets:

  • HCP 500 Subjects Release (S500): Used as a training set for all models.
  • BraTS2017: A test set containing high-grade and low-grade gliomas.
  • MS-SEG2015: A test set with MRI data from subjects experiencing sub-acute ischemic stroke at different times.

The paper does not introduce new benchmarks but evaluates AFSC on existing, well-established datasets for medical anomaly detection.

Key Insights and Ideas

  • Fourier Space as a Rich Source of Healthy Information: The paper posits that the amplitude and phase components of the Fourier transform of an image contain crucial information about its underlying healthy structure, which can be leveraged for anomaly detection.
  • Adaptive Compression for Generalization: By adaptively learning compression masks in the Fourier domain, AFSC can effectively filter out lesion-specific noise and retain essential healthy features, leading to better generalization to unseen anomalies.
  • Flexibility of AFSCm: The AFSCm module is designed to be plug-and-play, capable of enhancing existing pseudo-health generation-based AD methods by providing a more informative representation for the generator.
  • Sparsity as a Regularizer: The sparsity constraint on the learned masks encourages the model to focus on the most discriminative Fourier components, further improving efficiency and robustness.

Key Results

  • Improved Performance: AFSC consistently outperformed baseline methods (AnoGAN, AE, DRAEM) across various metrics (AUROCimg, APimg, DICE, AUROCpix, APpix) on both the BraTS2017 and MS-SEG2015 datasets.
  • Significant Gains with Integration: Integrating AFSCm with existing methods like AE and DRAEM led to substantial improvements in their performance, demonstrating the module's effectiveness. For instance, DRAEM_GW/AFSCm achieved the best results on the BraTS dataset for both DICE and APpix metrics.
  • Qualitative Improvements: Visualizations showed that AFSCm enabled generated pseudo-healthy images to be more anatomically accurate, making lesions appear more like healthy tissue and improving anomaly localization.
  • Validation of Adaptive Masks: Replacing the adaptive masks with fixed random masks of the same sparsity showed that the adaptively learned masks are indeed discriminative for critical health information.

Limitations and Future Work

  • Empirical Tuning of Sparsity: The current version of AFSC requires empirical tuning of the sparsity parameter, which could be a topic for future optimization. The paper suggests that smaller sparsity might be beneficial for more powerful generative models.
  • Further Exploration of Fourier Space: The authors believe that compressing information in Fourier space offers a novel perspective for anomaly detection, and further research into how to best utilize this information is a promising direction.
  • Cross-Dataset Testing: While AFSC performed well, the MS-SEG2015 dataset presented challenges due to small and subtle lesions, leading to generally unsatisfactory results for all methods, indicating the inherent difficulty of detecting such anomalies.

Summary updated

Abstract

Anomaly Detection (AD) on medical images enables a model to recognize any type of anomaly pattern without lesion-specific supervised learning. Data augmentation based methods construct pseudo-healthy images by "pasting" fake lesions on real healthy ones, and a network is trained to predict healthy images in a supervised manner. The lesion can be found by difference between the unhealthy input and pseudo-healthy output. However, using only manually designed fake lesions fail to approximate to irregular real lesions, hence limiting the model generalization. We assume by exploring the intrinsic data property within images, we can distinguish previously unseen lesions from healthy regions in an unhealthy image. In this study, we propose an Adaptive Fourier Space Compression (AFSC) module to distill healthy feature for AD. The compression of both magnitude and phase in frequency domain addresses the hyper intensity and diverse position of lesions. Experimental results on the BraTS and MS-SEG datasets demonstrate an AFSC baseline is able to produce promising detection results, and an AFSC module can be effectively embedded into existing AD methods.
BibTeX
@article{xu2022afsc,
  author = {Haote Xu and Yunlong Zhang and Liyan Sun and Chenxin Li and Yue Huang and Xinghao Ding},
  title = {{AFSC: Adaptive Fourier Space Compression for Anomaly Detection}},
  journal = {IEEE Trans. Industrial Informatics},
  year = {2022},
  doi = {10.1109/tii.2024.3423331},
}
Figure