A Systematic Survey of Fuzzy Deep Learning for Uncertain Medical Data

Yuanhang Zheng, Zeshui Xu, Tong Wu, Yi ZhangResearch Square 2023 · DOI

This systematic survey explores the burgeoning field of Fuzzy Deep Learning (FDL), a hybrid approach designed to handle the inherent messiness and uncertainty of medical data. By merging the human-like reasoning of fuzzy logic with the representational power of deep learning, researchers are creating more robust and interpretable tools for clinical decision-making.

TL;DR

Medical data is rarely "crisp"—it is riddled with noise, vague boundaries, and imprecise clinical notes. This paper surveys how Fuzzy Deep Learning bridges this gap, offering a framework to handle uncertainty that traditional deep learning models often ignore, ultimately leading to more reliable and interpretable medical AI.

What Problem Does This Paper Address?

Traditional Deep Learning (DL) operates on "crisp" values—binary logic where something is either $A$ or $B$. However, medicine lives in the gray areas. A tumor boundary in an ultrasound might be blurry; a doctor’s note might use vague descriptors; or a sensor might produce noisy data.

When standard DL models encounter this uncertainty, they often overfit to noise or fail to generalize, leading to "black box" decisions that clinicians find hard to trust. The authors address the need for a systematic categorization of how fuzzy set theory (which mathematically models "degrees of truth") can be integrated into deep architectures to handle this medical uncertainty.

Methods

The paper identifies four primary architectural frameworks for FDL:

  1. Sequential Framework: Fuzzy techniques (like Fuzzy C-means) are used for preprocessing, such as data normalization or initial image segmentation, before passing the results to a standard DL model (e.g., a CNN or LSTM).
  2. Simultaneous Framework: Fuzzy systems and DL models work in parallel on the same task. Their independent results are then fused to produce a final, more accurate diagnosis.
  3. Blending Framework: This is the most deeply integrated approach. Fuzzy logic or membership functions are embedded inside the neural network layers—for example, using a fuzzy membership function as an activation function.
  4. Integrated Framework: DL models are used first to extract features or generate probability maps, which are then processed by fuzzy techniques (like fuzzy integrals) to reach a final decision.

The survey covers various architectures, including Fuzzy Convolutional Neural Networks (FCNNs) for images and Fuzzy Long Short-Term Memory (FLSTM) for time-series data.

Datasets

As a systematic survey, this paper does not introduce a new dataset but reviews 513 articles that utilize a wide array of medical data sources:

  • Imaging: X-ray, CT, Ultrasound, and MRI datasets (e.g., for COVID-19 screening, lung pathology, and brain tumor detection).
  • Genomics: High-throughput data from The Cancer Genome Atlas (TCGA) for DNA-seq and RNA-seq analysis.
  • Electronic Health Records (EHR): Unstructured clinical nursing notes and mortality prediction data.
  • Traditional Chinese Medicine (TCM): Linguistic data used for constitutional type identification.

Key Insights and Ideas

  • The "Gray Area" Advantage: FDL excels at "boundary point processing." While standard models struggle with the fuzzy edges of an organ or lesion, fuzzy membership functions can accurately weight these uncertain pixels.
  • Interpretability: Because fuzzy logic is based on "if-then" rules that mimic human perception, FDL models are inherently more explainable than standard "black-box" neural networks.
  • Few-Shot Learning: Fuzzy logic provides a mathematical "prior" that helps models perform better when medical data is scarce—a common problem in rare disease research.
  • Multi-modal Fusion: FDL is particularly effective at combining heterogeneous data, such as merging low-frequency and high-frequency sub-images in MRIs or combining text records with imaging.

Key Results

The authors highlight a massive spike in research interest: the number of relevant publications grew from just 28 in 2018 to 172 in 2021.

  • Top Applications: Image segmentation, COVID-19 diagnosis, and breast cancer classification are the most dominant use cases.
  • Performance: The surveyed literature consistently shows that FDL models outperform "crisp" DL models in noisy environments, specifically in metrics like the Dice Similarity Coefficient (DSC) for segmentation and Area Under the Curve (AUC) for classification.

Limitations and Future Work

Despite the promise, the authors acknowledge several hurdles:

  • Subjectivity: Membership functions are often "preset" based on expert knowledge, which introduces human bias. Future work should focus on using DL to learn these fuzzy parameters automatically from data.
  • Computational Complexity: Adding fuzzy layers increases the number of parameters and the mathematical overhead, making models slower to train.
  • Outliers: Current FDL models still struggle with extreme outliers or "novelty" cases that fall entirely outside the learned fuzzy rules.

Future Directions: The authors point toward integrating FDL with Next-Generation Sequencing (NGS) for genomic medicine and applying it to the complex, linguistic-heavy domain of Traditional Chinese Medicine (TCM) health management.

Summary updated

Abstract

<title>Abstract</title> Intelligent medical industry is in a rapid stage of development around the world, followed by are the expanding market size and basic theories of intelligent medical diagnosis and decision-making. Deep learning models have achieved good practical results in medical domain. However, traditional deep learning is almost calculated and developed by crisp values, while imprecise, uncertain, and vague medical data is common in the process of diagnosis and treatment. Medical data is usually uncertain because of the presence of noise, artifact or high dimensional unstructured information in the data. Fortunately, fuzzy deep learning that originated from fuzzy sets, can effectively deal with uncertain and inaccurate information, providing new viewpoints for alleviating the problems above. Therefore, it is important and necessary to review the contributions of fuzzy deep learning for uncertain medical data in recent years. This paper first constructs four types of frameworks of fuzzy deep learning methods used for uncertain medical data, and investigates the status from three aspects: fuzzy deep learning methods, type of uncertain medical data and execution effects. Then the performance evaluation metrics of fuzzy deep learning models are analyzed in details. We have found that although fuzzy deep learning indeed develops the interpretability, low-quality images, boundary point processing, imprecise text records processing and multi-source heterogenous data fusion in medical domain, challenges still exist and future research directions are provided from the perspective of genetic data processing, intelligent medical decision-making based on complex cognition information, health management of Traditional Chinese Medical and few-shot learning.
BibTeX
@inproceedings{zheng2023systematic,
  author = {Yuanhang Zheng and Zeshui Xu and Tong Wu and Yi Zhang},
  title = {{A Systematic Survey of Fuzzy Deep Learning for Uncertain Medical Data}},
  booktitle = {Research Square},
  year = {2023},
  doi = {10.21203/rs.3.rs-3126621/v1},
}
Figure