English | EN 
Imagine training a deep learning model for medical imaging not on images, but on raw electromagnetic measurements. No pixels, no anatomical contours, only complex-valued signals describing how waves interact with tissue. This is the reality in microwave breast imaging.
The starting point is a collection of S-parameters, which encode how electromagnetic waves are reflected and transmitted across antennas and frequencies. These measurements capture subtle interactions between waves and biological tissue, including scattering, attenuation, and phase shifts.
However, these signals are not directly usable by machine learning models. Before any algorithm can learn from them, they must be transformed into structured, meaningful representations. This transformation raises a fundamental question: Should one design features explicitly, or let models learn them automatically?
From raw signals to structured representations
Microwave imaging systems generate inherently complex datasets. In multistatic configurations, measurements span multiple transmitting and receiving antennas, multiple frequencies, reflection and transmission coefficients, complex values (amplitude and phase).
This produces a high-dimensional tensor rather than a structured image. Unlike conventional imaging, there is no direct spatial grid. Instead, the data reflects how electromagnetic waves propagate and scatter through heterogeneous tissue. To make these data usable for learning algorithms, several representations are typically considered:
- Raw complex-valued tensors
- Separate magnitude and phase channels
- Frequency-domain signatures
- Time-domain signals (via inverse transforms)
- Reconstructed spatial maps using beamforming or tomography
Each choice encodes implicit assumptions about what aspects of the signal are most informative.
Feature Engineering: leveraging physical insight
Feature engineering involves transforming raw S-parameters into compact and informative descriptors before training a model. These features are often inspired by known physical principles, particularly the contrast in dielectric properties between different tissue types.
For example, malignant tissues often exhibit higher permittivity and conductivity, which leads to stronger reflections or altered propagation patterns.
Common engineered features include frequency-dependent energy or attenuation, phase differences across antenna pairs, signal asymmetries and contrast indicators, statistical summaries (mean, variance, entropy), spatial projections obtained from simple reconstruction methods. Feature engineering is particularly valuable when interpretability and robustness are priorities.
End-to-End Learning: learning directly from signals
End-to-end learning follows a different philosophy. Instead of manually defining features, models are trained directly on minimally processed inputs, such as raw S-parameter tensors or simple transformations thereof.
Deep learning architectures can automatically learn hierarchical representations that capture complex interactions between signals, frequencies, and antenna configurations. This is particularly relevant because microwave imaging involves nonlinear scattering effects, frequency-dependent responses and interactions influenced by geometry and tissue composition.
These factors contribute to the ill-posed nature of the inverse problem, where multiple internal configurations may produce similar measurements. While powerful, end-to-end approaches are often constrained by the limited availability of high-quality data.
| Approach | Strengths | Limitations |
| Feature Engineering | Efficient with limited data, a common constraint in experimental setupsInterpretable, as features can often be related to physical propertiesReduced dimensionality, lowering computational cost and overfitting risk | May discard subtle or complex patterns present in raw dataDepends on prior assumptions, which may not hold across datasetsLess adaptable to variations in acquisition modes or system configurations |
| End-to-End Learning | Preserves full information content of the signalsEnables discovery of non-obvious patternsScales well with larger and more diverse datasets | Requires large amounts of data to generalize effectivelyOften lacks interpretability, especially in clinical contextsSensitive to noise, calibration inconsistencies, and dataset biases |
Hybrid approaches: bridging physics and learning
In practice, the distinction between feature engineering and end-to-end learning is increasingly blurred. Many approaches combine both strategies to leverage their respective strengths. Examples include feeding physics-derived features into deep networks; combining reconstructed images with raw signal inputs; designing architectures that respect antenna geometry or acquisition structure; incorporating constraints derived from electromagnetic models. Such hybrid methods reflect a broader trend toward integrating domain knowledge and data-driven learning, rather than treating them as competing paradigms
Choosing the Right Representation
The way S-parameters are structured has direct implications for downstream tasks:
- Detection may rely on global signal patterns
- Localization may require spatially interpretable representations
- Size estimation may depend on frequency-dependent features
Moreover, datasets generated under different acquisition modes or experimental setups can introduce variability that affects model performance. Designing an effective pipeline therefore requires balancing preservation of physical information, model complexity, robustness to variability and interpretability of results.
Signals first, meaning later
Microwave imaging challenges conventional assumptions in medical AI. Instead of starting with images, it begins with signals that implicitly encode spatial information.
Transforming these signals into usable inputs is not a trivial preprocessing step, it is a core component of the learning pipeline. Feature engineering provides structure and interpretability, while end-to-end learning offers flexibility and discovery.
Rather than choosing one over the other, the most promising direction lies in combining both approaches, creating systems that are grounded in physics while capable of learning beyond predefined assumptions.
Ultimately, the goal is to convert complex electromagnetic measurements into meaningful, reliable insights, bridging the gap between raw signals and clinically relevant understanding.
Sources and further reading
- Khalid, N. et al., Emerging paradigms in microwave imaging technology for biomedical applications: unleashing the power of artificial intelligence, Nature / npj Imaging, 2024. https://www.nature.com/articles/s44303-024-00012-8
- Shao, W., Machine Learning in Microwave Medical Imaging and Lesion Detection, Diagnostics, 2025. https://www.mdpi.com/2075-4418/15/8/986
- Sharma, R.; Yurduseven, O., Quantitative Microwave Imaging using Deep Learning Network Guided by Plane Wave Equation, IEEE Transactions on Radar Systems, 2024. https://doi.org/10.1109/TRS.2024.3417519
- Origlia, C. et al., Review of Microwave Near-Field Sensing and Imaging Devices in Medical Applications, Sensors, 2024. https://www.mdpi.com/1424-8220/24/14/4515