THREE TASKS, THREE LEARNING STRATEGIES IN MICROWAVE BREAST IMAGING

In medical imaging, identifying a tumor is rarely a single problem. It is a sequence of interconnected tasks that progressively refine our understanding of what is happening inside the body. First, we ask: Is there a tumor? Then: Where is it? And finally: What are its properties?

In microwave imaging, this progression is especially important. Because we start from electromagnetic signals rather than direct images, extracting clinically meaningful information requires carefully designed learning strategies. Detection, localization, and characterization are not simply different outputs of the same model, they are fundamentally distinct problems, each with its own challenges and implications for artificial intelligence.

Detection: finding signals of abnormality

Detection is the most basic task: determining whether a tumor is present or not. From a machine learning perspective, this is typically framed as a binary classification problem. The input consists of structured microwave data, often derived from S-parameters, while the output is a probability indicating the presence of an anomaly. Because tumors create dielectric contrast with surrounding tissue, they alter scattering patterns in measurable ways.

Learning strategies for detection

Detection models often rely on:

• Global representations of the signal, capturing overall deviations
• Feature-based classifiers (e.g., SVM, random forests) when datasets are small
• Deep neural networks trained on raw or minimally processed data

Detection benefits from relatively simple learning strategies because it does not require precise spatial information. Instead, it leverages aggregated patterns across frequencies and antenna pairs.

The key challenges for detection are (i) subtle contrasts for small or early-stage tumors, (ii) high variability between breast compositions and (iii) sensitivity to noise and calibration errors. Despite these challenges, detection remains the most accessible entry point for AI in microwave imaging, often achieving promising results even with limited data.

Localization: mapping the tumor in space

Once a tumor is detected, the next question is its position. Localization aims to estimate the spatial coordinates of the anomaly, either in 2D or 3D. Unlike detection, localization requires models to infer spatial structure from signals that do not directly encode position. This makes the problem significantly more complex.

Learning strategies for localization

Localization methods typically involve:

• Image-based pipelines, where signals are first reconstructed into spatial maps (e.g., beamforming or tomography), then processed by CNNs
• Direct regression models predicting coordinates from structured signal inputs
• Heatmap-based approaches, estimating probability distributions over spatial grids

In multistatic systems, the diversity of signal paths can provide richer spatial cues, which are particularly useful for this task.

The key challenges for localization are (i) ill-posed mapping between signals and spatial positions, (ii) limited spatial resolution due to wavelength constraints and (iii) sensitivity to acquisition geometry and modelling assumptions. Localization shifts the problem from pattern recognition to spatial inference, requiring models to learn implicit representations of geometry and wave propagation.

Characterization: understanding tumor properties

Characterization goes beyond detection and localization. It aims to estimate properties of the detected tumor, such as size, shape, dielectric contrast, potential malignancy indicators, etc. This task is closer to quantitative imaging, where the goal is not just to detect an anomaly but to describe it in clinically meaningful terms.

Learning strategies for characterization

Characterization models often rely on:

• Regression approaches predicting continuous variables (e.g., tumor size)
• Multi-task learning, combining detection, localization, and property estimation
• Physics-informed models, linking signal behavior to dielectric properties

Because dielectric parameters vary with tissue composition and frequency, characterization often requires integrating multi-frequency information and leveraging physical priors.

The key challenges for characterization are (i) limited ground truth availability for quantitative properties, (ii) strong dependence on reconstruction accuracy and (iii) ambiguity between tumor size and contrast effects. Characterization is typically the most advanced and challenging task, as it requires both accurate spatial understanding and meaningful interpretation of signal variations.

Three tasks, three levels of complexity

Although detection, localization, and characterization are related, they differ in fundamental ways:

This hierarchy reflects increasing levels of difficulty:

• Detection focuses on global anomalies
• Localization requires spatial reasoning
• Characterization involves quantitative interpretation

Each step introduces new sources of uncertainty and requires more sophisticated models.

One model or multiple models?

A key design question is whether to train separate models for each task or build a single unified model that performs all tasks simultaneously. Separate models are easier to train and interpret, can be optimized independently and require coordination between stages whereas unified models enable shared representations, can exploit correlations between tasks and often require more data and careful design.

Recent trends favor multi-task learning, where a single model learns detection, localization, and characterization jointly. This approach reflects the interconnected nature of the tasks and can improve overall performance when data is sufficient.

The role of data and acquisition

The effectiveness of each task depends strongly on data quality and acquisition strategy. Detection can work with simpler datasets. Localization benefits from multistatic diversity. Characterization requires high-fidelity and calibrated measurements. Measurement variability, noise, and differences in acquisition setups can significantly affect all three tasks, reinforcing the importance of robust data pipelines.

From questions to clinical insight

Detection, localization, and characterization form a natural progression in microwave imaging. They transform raw electromagnetic measurements into increasingly refined answers to clinical questions: Is there something abnormal? Where is it? What is it like?

Each step demands a different perspective on the data, and therefore a different learning strategy. Rather than viewing these tasks as independent, the most recent approaches consider them as parts of a unified inference problem, where signals, spatial structure, and physical properties are learned together.

In this context, artificial intelligence is not only a tool for prediction. It becomes a framework for progressively extracting meaning from signals, bridging the gap between electromagnetic measurements and clinically relevant understanding.