Abstract
Microwave-based breast cancer detection has been proposed as a complementary approach to compensate for some drawbacks of existing breast cancer detection techniques. Among the existing microwave breast cancer detection methods, machine learning-type algorithms have recently become more popular. These focus on detecting the existence of breast tumours rather than performing imaging to identify the exact tumour position. A key step of the machine learning approaches is feature extraction. One of the most widely used feature extraction method is principle component analysis (PCA). However, it can be sensitive to signal misalignment. This paper presents an empirical mode decomposition (EMD)-based feature extraction method, which is more robust to the misalignment. Experimental results involving clinical data sets combined with numerically simulated tumour responses show that combined features from EMD and PCA improve the detection performance with an ensemble selection-based classifier.
Introduction
Breast cancer screening currently relies heavily on X-ray mammography and MRI. While effective, mammography involves ionizing radiation and uncomfortable breast compression, while MRI remains too expensive for widespread routine screening. Microwave-based detection offers a promising alternative: it is low-cost, non-ionizing, and utilizes the dielectric contrast between healthy and malignant tissues.
Recent research in this field has shifted from complex imaging reconstructions to binary detection tasks using machine learning. These systems aim to answer a simple "yes/no" question regarding the presence of a tumor. However, the success of these models depends entirely on the quality of the extracted features. Traditional methods like Principal Component Analysis (PCA) are highly sensitive to signal misalignment—often caused by hardware jitter or slight antenna movements in clinical settings—which can lead to significant drops in diagnostic accuracy.
Proposed Methodology
To address the sensitivity of global feature extraction methods, this paper proposes a hybrid pipeline that incorporates Empirical Mode Decomposition (EMD). EMD is a data-driven technique that decomposes non-stationary microwave signals into a set of Intrinsic Mode Functions (IMFs). Unlike Fourier transforms, EMD adapts to the local characteristics of the signal, making it inherently more robust to temporal shifts.
The proposed system processes signals from $M$ antenna pairs. For each channel, a 50-dimensional feature vector is constructed: 30 features are derived from traditional PCA to capture global variance, and 20 features are derived from EMD. The EMD features consist of four statistical markers—Mean Absolute Value ($\mu$), Standard Deviation ($\sigma$), Kurtosis ($\kappa$), and Mean Absolute Successive Difference ($S$)—calculated across the first five IMFs.
Robustness and Feature Stability
The primary advantage of the EMD-based approach is its shift-invariance. In clinical environments, electronic "jitter" can cause signals to arrive slightly earlier or later than expected. For a global method like PCA, this shift appears as a fundamental change in the data. In contrast, the statistical features derived from decomposed IMFs remain remarkably stable.
As shown in the feature analysis, the EMD-derived statistical markers (bottom plot) for a "shifted original" signal (green) closely track the "original" signal (blue). This stability ensures that the classifier can distinguish between a healthy signal with minor hardware noise and a true tumor response (yellow), which typically exhibits distinct local structural changes.
Experimental Results
The methodology was validated using a hybrid dataset consisting of clinical measurements from 12 healthy volunteers and numerically simulated tumor responses injected into the real-world data. The classification was performed using a cost-sensitive ensemble of Support Vector Machines (SVMs), optimized to minimize false negatives.
The results demonstrate a clear hierarchy in performance:
- PCA Features Only: AUC of 0.70
- EMD Features Only: AUC of 0.83
- Combined (EMD + PCA): AUC of 0.87
By combining the global perspective of PCA with the local, robust characteristics of EMD, the system achieved a significantly higher detection rate while maintaining a low false-alarm threshold. Specifically, the Mean Absolute Value and Standard Deviation of the IMFs were identified as the most critical features for successful detection.
Conclusion and Future Work
This study highlights the importance of shift-invariant feature extraction in microwave medical diagnostics. By integrating EMD into the machine learning pipeline, the researchers successfully mitigated the impact of signal misalignment, a common hurdle in clinical hardware implementation.
Future work will focus on validating this approach with a clinical population of patients with confirmed malignancies, as the current study relied on semi-synthetic tumor data. Additionally, improving signal-denoising techniques will be essential for maintaining performance in low Signal-to-Noise Ratio (SNR) scenarios where tumor responses are exceptionally weak.
Summary updated