Breast Cancer Detection Using Infrared Thermography: A Survey of Texture Analysis and Machine Learning Approaches

Larry Ryan, Sos С. AgaianBioengineering 2025 · DOI

TL;DR

This survey reviews the cutting-edge integration of texture analysis and machine learning for breast cancer detection using infrared thermography (IRT). It highlights IRT as a promising, cost-effective alternative to mammography, especially in resource-limited settings, and details the end-to-end pipeline from image processing to classification, identifying key challenges and future research directions.

What Problem Does This Paper Address?

Breast cancer remains a major global health concern, with early detection being crucial for improving survival rates. While mammography is the current gold standard, it has limitations like cost, accessibility barriers, discomfort, and radiation exposure, particularly for women with dense breasts. Infrared thermography (IRT) offers a radiation-free, non-invasive, and cost-effective alternative by detecting thermal signatures associated with tumors. However, IRT faces challenges such as limited clinical validation, lack of FDA approval as a primary screening tool, and difficulties in interpretation due to physiological variations and differing standards. This paper addresses the need for robust, automated systems to aid in IRT-based breast cancer detection by surveying the integration of texture analysis and machine learning, aiming to improve accuracy and reduce radiologist workload.

Methods

This survey provides a comprehensive review of the entire pipeline for computer-aided diagnosis (CAD) systems using IRT for breast cancer detection. The key stages covered are:

  • Image Acquisition: Discusses protocols and standards for capturing thermographic images.
  • Image Preprocessing: Covers techniques for noise reduction, contrast enhancement, normalization, and segmentation to isolate the breast region.
  • Feature Extraction: Details various methods for extracting discriminative information from thermographic images, categorized into:
  • Statistical Methods: First-order statistics (FOS), Co-occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM), Neighborhood Grey Tone Difference Matrix (NGTDM), and Tamura features.
  • Model-Based Methods: Fractal features (fractal dimension, Hurst exponent, lacunarity) and Vascular Network models.
  • Signal Processing Methods: Spatial domain filters (edge detection), Gabor filters, Wavelet analysis, Curvelet transform, and Histogram of Oriented Gradients (HOG).
  • Feature Reduction: Explores techniques to reduce the dimensionality of extracted features, including feature selection (t-test, Random Forest, NCA), dimension reduction (PCA, ICA), embedded methods (LASSO), and bio-inspired methods (Genetic Algorithm, PSO, Ant Colony Search, Firefly Algorithm).
  • Classification: Reviews various machine learning models used to classify images as normal or abnormal, with a focus on Support Vector Machines (SVMs) and neural networks.
  • Performance Assessment: Discusses metrics like accuracy, sensitivity, specificity, and AUC used to evaluate system performance.

The paper highlights that HOG and discrete wavelet-based features, often combined with SVM classifiers, have shown particularly strong performance.

Datasets

The paper discusses several datasets used in the reviewed studies:

  • DMR-IR Dataset: The most widely used public dataset, containing infrared images, mammograms, ROI masks, and clinical data for 293 patients. It's noted that images are not always classified per standard and may not clearly indicate left/right breast abnormalities.
  • HC-UFPE Dataset: Used in several studies, often converted from RGB to grayscale.
  • Ann Arbor Thermography Dataset: Used in combination with DMR-IR in some studies.
  • Mendeley Data: A dataset released in 2024, noted as being compliant with the American Standard.
  • DBT-TU-JU Dataset: A dataset used for breast thermogram analysis.
  • Private Datasets: Many studies utilize private datasets, which are not publicly available, making direct comparison difficult.

A significant challenge identified is the lack of standardized, publicly available datasets with comprehensive annotations (e.g., tumor location, type).

Key Insights and Ideas

  • IRT as a Viable Alternative: The survey strongly supports IRT, especially when combined with texture analysis and machine learning, as a cost-effective, non-invasive, and radiation-free alternative to mammography, particularly for specific populations (dense breasts, young women, pregnant/lactating women) and in resource-limited regions.
  • Texture Analysis is Key: The paper emphasizes that texture features are crucial for capturing subtle patterns in thermographic images that indicate cancerous lesions, often outperforming raw pixel data or simpler statistical measures.
  • HOG and Wavelets Shine: Histogram of Oriented Gradients (HOG) and discrete wavelet-based features consistently emerge as top-performing feature extraction methods, demonstrating robustness to illumination variations and geometric transformations.
  • SVM Dominance: Support Vector Machines (SVMs) are frequently the classifier of choice, often achieving the best results across various studies, though deep learning models are also showing promise.
  • Data Limitations are Critical: The heavy reliance on limited, often private, datasets like DMR-IR is a major bottleneck, hindering reproducibility and generalization. The lack of standardized protocols and annotations further complicates comparisons.
  • Explainability is Lacking: While CAD systems show high performance, their "black box" nature is a significant limitation for clinical adoption. Improving the explainability of these models for radiologists is a critical area for future research.

Key Results

  • Many studies report high performance metrics, with accuracies often exceeding 90%, and sensitivities and specificities in the high 90s.
  • For example, one study achieved 99.17% accuracy, 100% sensitivity, and 98% specificity using DWNN features fed into an SVM classifier.
  • Another study using HOG features achieved 98.27% accuracy and 98.15% F1-score.
  • The paper notes that while clinical trials are limited, thermography's performance is steadily increasing and comparable to mammography.

Limitations and Future Work

Limitations:

  • Limited Clinical Validation: Many studies are based on limited datasets, leading to high variability and potentially inaccurate results.
  • Over-reliance on DMR-IR: The DMR-IR dataset is heavily used, but it has limitations in terms of standardization and annotation.
  • Lack of Ablation Studies: The impact of specific components, like segmentation, on overall performance is often not rigorously tested.
  • Inconsistent Reporting: Performance metrics and experimental setups vary significantly across studies, making direct comparisons difficult.
  • Physiological Factors Ignored: Most models do not account for physiological factors (e.g., menstrual cycle, hormonal therapy) that can influence thermographic patterns.
  • Lack of Explainability: The "black box" nature of many CAD systems hinders radiologist trust and understanding.

Future Work:

  • Standardized Benchmark Dataset: Development of a large, unified benchmark dataset with standardized imaging protocols and comprehensive annotations is crucial.
  • Ablation Studies: Rigorous testing of individual components (e.g., segmentation, feature reduction) to understand their contribution to performance.
  • Improved Clinical Validation: Conducting more extensive clinical trials to validate the effectiveness of IRT-based CAD systems.
  • Multimodal Approaches: Integrating IRT with other imaging modalities (mammography, ultrasound, MRI) and clinical data for enhanced diagnostic accuracy.
  • Explainability and Interpretability: Developing methods to make CAD system recommendations understandable to radiologists, potentially through visual representations or feature importance analysis.
  • Addressing Physiological Variations: Incorporating models that account for factors influencing thermographic patterns.
  • Federated Learning: Exploring collaborative approaches to build models on diverse patient data while protecting privacy.
  • Wearable Devices: Investigating the potential of temperature sensor-based techniques integrated into wearable devices for remote breast cancer screening.

Summary updated

Abstract

Breast cancer remains a leading cause of cancer-related deaths among women worldwide, highlighting the urgent need for early detection. While mammography is the gold standard, it faces cost and accessibility barriers in resource-limited areas. Infrared thermography is a promising cost-effective, non-invasive, painless, and radiation-free alternative that detects tumors by measuring their thermal signatures through thermal infrared radiation. However, challenges persist, including limited clinical validation, lack of Food and Drug Administration (FDA) approval as a primary screening tool, physiological variations among individuals, differing interpretation standards, and a shortage of specialized radiologists. This survey uniquely focuses on integrating texture analysis and machine learning within infrared thermography for breast cancer detection, addressing the existing literature gaps, and noting that this approach achieves high-ranking results. It comprehensively reviews the entire processing pipeline, from image preprocessing and feature extraction to classification and performance assessment. The survey critically analyzes the current limitations, including over-reliance on limited datasets like DMR-IR. By exploring recent advancements, this work aims to reduce radiologists' workload, enhance diagnostic accuracy, and identify key future research directions in this evolving field.
BibTeX
@inproceedings{ryan2025breast,
  author = {Larry Ryan and Sos С. Agaian},
  title = {{Breast Cancer Detection Using Infrared Thermography: A Survey of Texture Analysis and Machine Learning Approaches}},
  booktitle = {Bioengineering},
  year = {2025},
  doi = {10.3390/bioengineering12060639},
}
Figure