Journal Home
Login
Register
Open Access Journal
Submit a Paper
Propose a Special lssue
Open Access
ARTICLE
The Resilience of Deep Learning Models for Breast Cancer Detection: A Quantitative Analysis of Performance Under Diverse Noise Conditions in Thermal Imaging
Issue Vol. 2 No. 01 (2025): Volume 02 Issue 01 --- Section Articles --- Published Date: 2025-01-06
Abstract
Breast cancer continues to be a leading cause of mortality among women globally, making early and accurate detection paramount. Infrared thermography has emerged as a promising non-invasive, radiation-free diagnostic modality that identifies potential malignancies by capturing the subtle temperature variations associated with tumor metabolism and angiogenesis. The integration of deep learning, particularly Convolutional Neural Networks (CNNs), has significantly advanced the analytical power of thermography. However, the inherent susceptibility of thermal images to various types of electronic and environmental noise presents a critical challenge to the reliability of these automated systems. The performance of deep learning models under realistic, noisy conditions is not yet fully understood.
This study provides a comprehensive and systematic evaluation of a state-of-the-art, modified Inception-based deep learning model's robustness to noise in the context of early breast cancer detection. We quantified the model's diagnostic performance on a large dataset of thermal images systematically corrupted by four distinct and clinically relevant noise types: Gaussian, speckle, salt-and-pepper, and Poisson noise. The intensity of each noise was varied across a wide range to identify performance degradation profiles and critical "tipping points."
Our results demonstrate that while the model achieves exceptional accuracy (99.975%) on clean, noise-free images, its performance is significantly impacted by noise. The nature and severity of this degradation are highly dependent on the noise type. Impulsive noise, such as salt-and-pepper, caused a drastic decline in accuracy to 51.58% at a density of 0.3. Similarly, high-variance speckle noise reduced accuracy to 43.86%. In contrast, the model exhibited greater resilience to Gaussian and Poisson noise, maintaining high accuracy across most tested intensities. Crucially, the application of a pre-processing denoising filter was shown to be highly effective, restoring classification accuracy on corrupted images to near-perfect levels (>99.9%).
This research underscores the critical vulnerability of deep learning-based diagnostic systems to image noise and establishes quantitative benchmarks for model robustness. The findings highlight the indispensable role of advanced noise mitigation strategies and rigorous imaging protocols in developing reliable and clinically viable AI-powered tools for breast cancer thermography
Keywords
References
[1] Hanf V, Kreienberg R. Corpus Uteri. 2020.
[2] Bini SA. Artificial Intelligence Machine Learning Deep Learning and Cognitive Computing: What Do These Terms Mean and How Will They Impact Health Care? J Arthroplasty. 2018;33:2358-2361.
[3] Yadav P, Jethani V. Breast Thermograms Analysisfor Cancer Detection Using Feature Extraction and Data Mining Technique. ACM Int Conf Proceeding Ser. 2016:1-5.
[4] Din NM, Dar RA, Rasool M, Assad A. Breast Cancer Detection Using Deep Learning: Datasets Methods and Challenges Ahead. Comput Biol Med. 2022;149:106073.
[5] Salvi S, Kadam A. Breast Cancer Detection Using Deep Learning and IoT Technologies. J Phys Conf Ser. 2021;1831:012030.
pdf
