⚡ electrical engineering

EfficientNet-B5 Based Breast Tumor Classification from CWT Morlet Scalogram Images of Microwave S-Parameters

This paper proposes an energy-guided signal-level augmentation method to overcome data scarcity in microwave breast imaging, achieving 94.7% accuracy in classifying benign and malignant tumors by converting augmented S-parameter responses into CWT Morlet scalograms and fine-tuning an EfficientNet-B5 classifier.

Original authors: Avik Majumder, Edison Kho, Banani Basu, Arnab Nandi

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Avik Majumder, Edison Kho, Banani Basu, Arnab Nandi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Breast cancer remains one of the most significant health challenges for women worldwide, and finding it early is often the difference between life and death. For decades, doctors have relied on tools like mammograms, which use X-rays, or ultrasound, which uses sound waves, to look inside the breast. While these methods are effective, they come with drawbacks: X-rays involve radiation and can be uncomfortable because the breast must be squeezed flat, while ultrasound results can vary depending on the skill of the technician. In recent years, scientists have turned their attention to a different kind of wave: microwaves. Unlike X-rays, microwaves are a form of non-ionizing radiation, meaning they do not carry the same risks of damaging DNA, and they do not require squeezing the breast. This technique works because cancerous tissue behaves differently than healthy tissue when hit by these waves; the tumor absorbs and reflects the energy in a unique way. However, teaching a computer to recognize these subtle differences has been difficult because there are very few real-world examples of microwave signals from breast tumors to learn from.

A team of researchers at the National Institute of Technology Silchar in India has developed a new way to solve this data shortage, allowing them to train powerful computer programs to distinguish between harmless and dangerous breast tumors using microwave signals. Instead of trying to find more patients or build more expensive equipment, the researchers created a method to intelligently multiply the existing data they had. They started by measuring how microwave signals bounced off and passed through a model of a human breast, known as a phantom, which was designed to mimic the electrical properties of real tissue. This model included both healthy tissue and simulated tumors, some benign and some malignant. Using a pair of small antennas, they recorded the signals as they moved around the model, capturing thousands of data points that represented the unique "fingerprint" of the breast's internal structure.

The challenge was that these raw signals were too few to teach a sophisticated computer system effectively. If the researchers simply copied the data or made random changes to it, the computer would learn the wrong patterns or miss the subtle clues that indicate a tumor. To fix this, the team devised a strategy based on the energy within the signals. They realized that not every part of a signal was equally important; some sections contained the strongest information about the tumor, while others were just background noise. They broke the signals into pieces based on how much energy each part held, rather than just cutting them into equal time slices. Then, they carefully stitched together the most informative pieces from different signals of the same type—mixing benign pieces with other benign pieces, and malignant with malignant. To ensure the new signals didn't sound jarring or artificial, they smoothed out the connections between the pieces. This process created thousands of new, realistic training examples from a small original set, effectively teaching the computer what to look for without inventing fake data.

Once they had this expanded library of signals, the researchers converted them into images that a computer could easily understand. They transformed the raw wave data into visual maps called scalograms, which show how the signal's frequency and strength change over time, much like a musical score shows notes changing over time. These images were then fed into a deep learning model, a type of artificial intelligence designed to recognize patterns in pictures. The researchers tested several different models, but they found that one specific architecture, known as EfficientNet-B5, performed the best. This model was able to learn the complex visual patterns in the microwave images that distinguished a benign growth from a malignant one.

The results of this approach were impressive. When tested on data it had never seen before, the system correctly identified the type of tumor in nearly 95 percent of cases. This accuracy was higher than other popular computer models tested in the same study, and it outperformed several recent studies that used different imaging methods like ultrasound or thermal cameras. The researchers also checked how stable their system was by testing it multiple times with different splits of the data, and the results remained consistently high, with very little variation. They found that the more they used their energy-based method to expand the data, the better the computer became at making the right call, up to a certain point where adding more data offered diminishing returns.

This work suggests that microwave imaging, when paired with smart data techniques, could become a viable, comfortable, and safe alternative for screening breast cancer. The method does not require the patient to be exposed to radiation or subjected to physical compression, making it a potentially more patient-friendly option. While the study was conducted using a model of a breast rather than real patients, the high level of accuracy achieved in these controlled experiments provides a strong foundation for future development. The researchers demonstrated that by understanding the physics of the signals and using that knowledge to generate better training data, it is possible to overcome the limitations of small datasets and build reliable tools for early cancer detection.

Technical Summary: EfficientNet-B5 Based Breast Tumor Classification from CWT Morlet Scalogram Images of Microwave S-Parameters

Problem Statement
Microwave breast imaging (MBI) offers a non-ionizing, non-compressive alternative for breast tumor detection by leveraging the dielectric contrast between healthy and malignant tissues. However, the application of deep learning in this domain is significantly hindered by the scarcity of experimentally measured samples. Acquiring microwave data requires time-consuming fabrication of breast phantoms and extensive scanning procedures. Consequently, deep learning models trained on such limited datasets are prone to overfitting and poor generalization. Furthermore, standard data augmentation techniques used in computer vision (e.g., rotation, flipping) or generic signal processing (e.g., random segmentation) are often unsuitable for microwave S-parameters, as they can distort the physical meaning of the time-frequency data or introduce artificial phase discontinuities.

Methodology
The proposed framework addresses these limitations through a physics-aware data augmentation strategy and a deep learning classification pipeline:

Data Acquisition and Preprocessing:
- Hardware: A compact UWB microstrip antenna pair operating at 5.691 GHz was designed and used to measure reflection ( $S_{11}$ ) and transmission ( $S_{21}$ ) parameters from a half-spherical breast phantom. The phantom mimics the dielectric properties (conductivity and permittivity) of human skin, fat, glandular tissue, and tumors (benign and malignant) using the Cole-Cole dispersion relation.
- Measurement: Signals were captured over a 2–12 GHz frequency range with the antenna pair rotated around the phantom in 1° steps, yielding 360 pairs of S-parameters for each class (benign, malignant, and normal).
- Signal Processing: Frequency-domain responses were converted to the time domain using a rational fitting model followed by an inverse Fourier transform. A clutter-removal procedure (subtracting tumor-free responses and averaging background terms) was applied to suppress unwanted reflections. Finally, wavelet-based denoising (db4 wavelet, level-6 soft thresholding) was identified as the most effective method to reduce noise while preserving signal envelopes and spectral characteristics.
Energy-Guided Signal Augmentation:
- To expand the limited dataset without violating physical constraints, the authors proposed an energy-guided segmentation method. Unlike equal-length segmentation, signals were divided into segments containing nearly equal amounts of electromagnetic energy.
- High-energy segments, which contain stronger scattering information from the tumor region, were selected and recombined to generate new synthetic signals.
- Smooth transitions (interpolation and crossfading) were applied at segment boundaries to minimize artificial discontinuities. This process generated 3,600 augmented samples per class for both $S_{11}$ and $S_{21}$ , expanding the dataset from 720 original samples to 7,200 total samples.
- t-SNE visualization confirmed that the augmented samples expanded the feature space coverage while maintaining the clustering patterns of the original data.
Scalogram Generation and Classification:
- The augmented time-domain signals were transformed into 2D images using the Continuous Wavelet Transform (CWT) with a Morlet mother wavelet. The study utilized the magnitude of the CWT coefficients ( $|W_x(a,b)|$ ) rather than the squared magnitude to preserve a balanced dynamic range without overemphasizing dominant responses.
- These scalogram images were resized to $456 \times 456$ pixels and fed into a modified ImageNet-pretrained EfficientNet-B5 model.
- The classification head was customized with global average pooling, two dropout layers (rate 0.55), and fully connected layers. The model was trained using a two-stage transfer learning approach: freezing the backbone for the first two epochs, followed by joint fine-tuning. Stochastic depth and label smoothing were employed to prevent overfitting.

Key Contributions

Energy-Guided Augmentation: The paper introduces a novel signal-level augmentation technique that respects the physical energy distribution of microwave responses. This method increases dataset size while preserving the scattering behavior essential for tumor detection, avoiding the pitfalls of random or geometric signal manipulation.
Optimized Deep Learning Architecture: The study systematically evaluates various EfficientNet variants (B0–B7) and other architectures (DenseNet-201, InceptionV3, Xception). It identifies EfficientNet-B5 as the optimal backbone, offering the best balance between feature extraction capability and computational cost for this specific dataset size and complexity.
Robust Denoising and Transformation: The integration of wavelet denoising and CWT-based scalogram generation creates a robust pipeline for converting raw microwave S-parameters into discriminative image features suitable for deep learning.

Results
The proposed framework was evaluated using 5-fold cross-validation on the augmented dataset:

Performance: The EfficientNet-B5 model achieved a mean accuracy of 94.78% ± 1.91%, with a precision of 97.85%, recall of 91.61%, and an F1-score of 94.59%.
Discriminative Power: The model demonstrated a mean Area Under the Curve (AUC) of 0.9920 ± 0.0046, indicating excellent separation between benign and malignant classes.
Comparison: The proposed method outperformed other EfficientNet variants (where B6 and B7 showed slight performance declines likely due to overfitting on the limited data) and other state-of-the-art models like DenseNet-201 and Xception.
Ablation: Experiments showed that combining signal-level (energy-guided) and image-level augmentations yielded the most stable training curves and best generalization compared to using either method alone or no augmentation.

Significance and Claims
The paper claims that the proposed energy-guided augmentation strategy effectively mitigates the data scarcity problem in experimental microwave breast imaging. By generating synthetic signals that retain the physical scattering characteristics of the original measurements, the method enables deep learning models to learn stable, tumor-related features from limited experimental data. The authors emphasize that the framework provides a practical balance between high classification performance and feasibility, offering a non-ionizing, non-compressive screening tool that does not require the extensive computational resources of iterative reconstruction techniques. The results suggest that this approach is a viable step toward portable and cost-effective microwave-based breast tumor detection systems.

More like this