Diagnosing Breast Cancer with Microwave Technology: remaining challenges and potential solutions with machine learning

Machine learning and new insights for breast cancer diagnosis

Breast cancer (BC) is the most prominent form of cancer among females all over the world. The current methods of BC detection include X-ray mammography, ultrasound, computed tomography, magnetic resonance imaging, positron emission tomography and breast thermographic techniques. More recently, machine learning (ML) tools have been increasingly employed in diagnostic medicine for its high efficiency in detection and intervention. The subsequent imaging features and mathematical analyses can then be used to generate ML models, which stratify, differentiate and detect benign and malignant breast lesions. Given its marked advantages, radiomics is a frequently used tool in recent research and clinics. Artificial neural networks and deep learning (DL) are novel forms of ML that evaluate data using computer simulation of the human brain. DL directly processes unstructured information, such as images, sounds and language, and performs precise clinical image stratification, medical record analyses and tumour diagnosis. Herein, this review thoroughly summarizes prior investigations on the application of medical images for the detection and intervention of BC using radiomics, namely DL and ML. The aim was to provide guidance to scientists regarding the use of artificial intelligence and ML in research and the clinic.

Dielectric Characterization of Healthy Human Teeth from 0.5 to 18 GHz with an Open-Ended Coaxial Probe

Dental caries is a major oral health issue which compromises oral health, as it is the main cause of oral pain and tooth loss. Early caries detection is essential for effective clinical intervention. However, methods commonly employed for its diagnosis often fail to detect early caries lesions, which motivates the research for more effective diagnostic solutions. In this work, the relative permittivity of healthy permanent teeth, in caries-prone areas, was studied between 0.5 and 18 GHz. The reliability of such measurements is an important first step to, ultimately, evaluate the feasibility of a microwave device for caries detection. The open-ended coaxial probe technique was employed. Its performance showed to be compromised by the poor probe-tooth contact. We proposed a method based on applying coupling media to reduce this limitation. A decrease in the measured relative permittivity variability was observed when the space between the probe tip and tooth surface was filled by coupling media instead of air. The influence of the experimental conditions in the measurement result was found to be less than 5%. Measurements conducted in ex vivo teeth showed that the relative permittivity of the dental crown and root ranges between 10.0-11.0 and 8.0-9.5, respectively.

Evaluating the Performance of Algorithms in Axillary Microwave Imaging towards Improved Breast Cancer Staging

Breast cancer is the most common and the fifth deadliest cancer worldwide. In more advanced stages of cancer, cancer cells metastasize through lymphatic and blood vessels. Currently there is no satisfactory neoadjuvant (i.e., preoperative) diagnosis to assess whether cancer has spread to neighboring Axillary Lymph Nodes (ALN). This paper addresses the use of radar Microwave Imaging (MWI) to detect and determine whether ALNs have been metastasized, presenting an analysis of the performance of different artifact removal and beamformer algorithms in distinct anatomical scenarios. We assess distinct axillary region models and the effect of varying the shape of the skin, muscle and subcutaneous adipose tissue layers on single ALN detection. We also study multiple ALN detection and contrast between healthy and metastasized ALNs. We propose a new beamformer algorithm denominated Channel-Ranked Delay-Multiply-And-Sum (CR-DMAS), which allows the successful detection of ALNs in order to achieve better Signal-to-Clutter Ratio, e.g., with the muscle layer up to 3.07 dB, a Signal-to-Mean Ratio of up to 20.78 dB and a Location Error of 1.58 mm. In multiple target detection, CR-DMAS outperformed other well established beamformers used in the context of breast MWI. Overall, this work provides new insights into the performance of algorithms in axillary MWI.

Antenna Excitation Optimization with Deep Learning for Microwave Breast Cancer Hyperthermia

Microwave hyperthermia (MH) requires the effective calibration of antenna excitations for the selective focusing of the microwave energy on the target region, with a nominal effect on the surrounding tissue. To this end, many different antenna calibration methods, such as optimization techniques and look-up tables, have been proposed in the literature. These optimization procedures, however, do not consider the whole nature of the electric field, which is a complex vector field; instead, it is simplified to a real and scalar field component. Furthermore, most of the approaches in the literature are system-specific, limiting the applicability of the proposed methods to specific configurations. In this paper, we propose an antenna excitation optimization scheme applicable to a variety of configurations and present the results of a convolutional neural network (CNN)-based approach for two different configurations. The data set for CNN training is collected by superposing the information obtained from individual antenna elements. The results of the CNN models outperform the look-up table results. The proposed approach is promising, as the phase-only optimization and phase-power-combined optimization show a 27% and 4% lower hotspot-to-target energy ratio, respectively, than the look-up table results for the linear MH applicator. The proposed deep-learning-based optimization technique can be utilized as a protocol to be applied on any MH applicator for the optimization of the antenna excitations, as well as for a comparison of MH applicators.

Heterogeneous Skin Phantoms for Experimental Validation of Microwave-Based Diagnostic Tools

Considerable exploration has been done in recent years to exploit the reported inherent dielectric contrast between healthy and malignant tissues for a range of medical applications. In particular, microwave technologies have been investigated towards new diagnostic medical tools. To assess the performance and detection capabilities of such systems, tissue-mimicking phantoms are designed for controlled laboratory experiments. We here report phantoms developed to dielectrically represent malign skin lesions such as liposarcoma and nonsyndromic multiple basal cell carcinoma. Further, in order to provide a range of anatomically realistic scenarios, and provide meaningful comparison between different phantoms, cancer-mimicking lesions are inserted into two different types of skin phantoms with varying tumor–skin geometries. These configurations were measured with a microwave dielectric probe (0.5–26.5 GHz), yielding insight into factors that could affect the performance of diagnostic and detection tools.

Radial Basis Function for Breast Lesion Detection from MammoWave Clinical Data

Recently, a novel microwave apparatus for breast lesion detection (MammoWave), uniquely able to function in air with 2 antennas rotating in the azimuth plane and operating within the band 1-9 GHz has been developed. Machine learning (ML) has been implemented to understand information from the frequency spectrum collected through MammoWave in response to the stimulus, segregating breasts with and without lesions. The study comprises 61 breasts (from 35 patients), each one with the correspondent output of the radiologist's conclusion (i.e., gold standard) obtained from echography and/or mammography and/or MRI, plus pathology or 1-year clinical follow-up when required. The MammoWave examinations are performed, recording the frequency spectrum, where the magnitudes show substantial discrepancy and reveals dissimilar behaviours when reflected from tissues with/without lesions. Principal component analysis is implemented to extract the unique quantitative response from the frequency response for automated breast lesion identification, engaging the support vector machine (SVM) with a radial basis function kernel. In-vivo feasibility validation (now ended) of MammoWave was approved in 2015 by the Ethical Committee of Umbria, Italy (N. 6845/15/AV/DM of 14 October 2015, N. 10352/17/NCAV of 16 March 2017, N 13203/18/NCAV of 17 April 2018). Here, we used a set of 35 patients. According to the radiologists conclusions, 25 breasts without lesions and 36 breasts with lesions underwent a MammoWave examination. The proposed SVM model achieved the accuracy, sensitivity, and specificity of 91%, 84.40%, and 97.20%. The proposed ML augmented MammoWave can identify breast lesions with high accuracy.

Multi-frequency symmetry difference electrical impedance tomography with machine learning for human stroke diagnosis

OBJECTIVE

Multi-frequency symmetry difference electrical impedance tomography (MFSD-EIT) can robustly detect and identify unilateral perturbations in symmetric scenes. Here, an investigation is performed to assess if the algorithm can be successfully applied to identify the aetiology of stroke with the aid of machine learning.

METHODS

Anatomically realistic four-layer finite element method models of the head based on stroke patient images are developed and used to generate EIT data over a 5 Hz-100 Hz frequency range with and without bleed and clot lesions present. Reconstruction generates conductivity maps of each head at each frequency. Application of a quantitative metric assessing changes in symmetry across the sagittal plane of the reconstructed image and over the frequency range allows lesion detection and identification. The algorithm is applied to both simulated and human (n = 34 subjects) data. A classification algorithm is applied to the metric value in order to differentiate between normal, haemorrhage and clot values.

MAIN RESULTS

An average accuracy of 85% is achieved when MFSD-EIT with support vector machines (SVM) classification is used to identify and differentiate bleed from clot in human data, with 77% accuracy when differentiating normal from stroke in human data.

CONCLUSION

Applying a classification algorithm to metrics derived from MFSD-EIT images is a novel and promising technique for detection and identification of perturbations in static scenes.

SIGNIFICANCE

The MFSD-EIT algorithm used with machine learning gives promising results of lesion detection and identification in challenging conditions like stroke. The results imply feasible translation to human patients.

Classification of breast tumor models with a prototype microwave imaging system

PURPOSE

The assessment of the size and shape of breast tumors is of utter importance to the correct diagnosis and staging of breast cancer. In this paper, we classify breast tumor models of varying sizes and shapes using signals collected with a monostatic ultra-wideband radar microwave imaging prototype system with machine learning algorithms specifically tailored to the collected data.

METHODS

A database comprising 13 benign and 13 malignant tumor models with sizes between 13 and 40 mm was created using dielectrically representative tissue mimicking materials. These tumor models were placed inside two breast phantoms: a homogeneous breast phantom and a breast phantom with clusters of fibroglandular mimicking tissue, accounting for breast heterogeneity. The breast phantoms with tumors were imaged with a monostatic microwave imaging prototype system, over a 1-6 GHz frequency range. The classification of benign and malignant tumors embedded in the two breast phantoms was completed, and tumor classification was evaluated with Principal Component Analysis as a feature extraction method, and tuned Naïve Bayes (NB), decision trees (DT), and k-nearest neighbours (kNN) as classifiers. We further study which antenna positions are better placed to classify tumors, discuss the feature extraction method and optimize classification algorithms, by tuning their hyperparameters, to improve sensitivity, specificity and the receiver operating characteristic curve, while ensuring maximum generalization and avoiding overfitting and data contamination. We also added a realistic synthetic skin response to the collected signals and examined its global effect on classification of benign vs malignant tumors.

RESULTS

In terms of global classification performance, kNN outperformed DT and NB machine learning classifiers, achieving a classification accuracy of 96.2% when classifying between benign and malignant tumor phantoms in a homogeneous breast phantom (both when the skin artifact is and is not considered).

CONCLUSIONS

We experimentally classified tumor models as benign or malignant with a microwave imaging system, and we showed a methodology that can potentially assess the shape of breast tumors, which will give further insight into the correct diagnosis and staging of breast cancer.

3D Simulations of Intracerebral Hemorrhage Detection Using Broadband Microwave Technology

Early, preferably prehospital, detection of intracranial bleeding after trauma or stroke would dramatically improve the acute care of these large patient groups. In this paper, we use simulated microwave transmission data to investigate the performance of a machine learning classification algorithm based on subspace distances for the detection of intracranial bleeding. A computational model, consisting of realistic human head models of patients with bleeding, as well as healthy subjects, was inserted in an antenna array model. The Finite-Difference Time-Domain (FDTD) method was then used to generate simulated transmission coefficients between all possible combinations of antenna pairs. These transmission data were used both to train and evaluate the performance of the classification algorithm and to investigate its ability to distinguish patients with versus without intracranial bleeding. We studied how classification results were affected by the number of healthy subjects and patients used to train the algorithm, and in particular, we were interested in investigating how many samples were needed in the training dataset to obtain classification results better than chance. Our results indicated that at least 200 subjects, i.e., 100 each of the healthy subjects and bleeding patients, were needed to obtain classification results consistently better than chance (p < 0.05 using Student's t-test). The results also showed that classification results improved with the number of subjects in the training data. With a sample size that approached 1000 subjects, classifications results characterized as area under the receiver operating curve (AUC) approached 1.0, indicating very high sensitivity and specificity.