Skip to main content

Microwave Diagnosis of Breast Cancer with Open Ended Contact Probes

Periodic Reporting for period 1 - MIDxPRO (Microwave Diagnosis of Breast Cancer with Open Ended Contact Probes)

Période du rapport: 2017-05-08 au 2019-05-07

Diagnosis and treatment of breast cancer are currently performed in several steps. During the diagnosis, the tumor tissues are imaged with either mammography or ultrasound. Following the initial identification of the tumor, the malignancy is determined with biopsy. Similarly, during cancer resection surgeries, the excised sample should be analyzed by a pathology specialist to ensure complete removal of the malignant tissue, named Frozen Section Analysis (FSA). Both FSA and biopsy requires a specialist, time-consuming, and costly. The failure of the pathology in diagnosis may cause the growth and spread of the malignant tumor. Pathological failure during resection surgeries may cause the recurrence of the disease. In fact, many breast cancer patients go through radical mastectomy, which entails complete removal of the breast, to minimize the recurrence risk despite several severe complications.

One technique that can emerge as an alternative to the pathology during diagnosis and treatment is the open-ended contact probe method. This technique is currently available for laboratory use. Due to the inherent dielectric property discrepancy between the malignant and benign tissues, the application of this technique as a biopsy device was envisioned in the literature. However, the biopsy application of the probe was overlooked due to the high measurement error rate of the technique.

The improvement in the diagnosis and treatment techniques for breast cancer can aid the correct/early diagnosis of the disease. Similarly, enabling an automated decision-making system that can characterize the tissue type would both decrease the human error and the cost of analysis. Such a device can be deployed to rural areas to determine the malignancies in the physician’s office.

The main goal of the project is to enable the practical usage of open-ended contact probe technique by decreasing the dielectric property measurement error. This can be done by addressing the error sources of the system which includes,
• introducing a new mathematical approach,
• optimizing the probe structure,
• adopting Machine Learning algorithms,
• verifying the modifications with in vivo experiments.
Commercial probes are reporting an error rate of 5% in a laboratory environment while in the literature measurement error reported as high as 30% for a practical setting. By implementing the above-listed objectives, we can address the error sources while enabling the practical application of the technique.
To address the error due to the solution of the mathematical formulation, we approached the mathematical formulation as an inverse problem. The conventional approach includes solving the problem iteratively which does not guarantee the convergence to an optimal solution. To address the frequency dependence, the Cole-Cole parameters were derived instead of dielectric properties and the mathematical formulation is solved with a Newton-based method with Tikhonov regularization. The algorithm is tested with two benchmark liquids and the results agreed well with both the results of commercial technique and the literature.

Probes with three different apertures were simulated, fabricated, and tested. The probes were tested with benchmark liquids. The measurements revealed that the dielectric properties obtained with the 0.5 mm probe were close to the dielectric properties collected with the 2.2 mm probe with a good agreement with literature. Also, to understand the probe sensing depth, we performed sensing depth measurements with the commercial 2.2 mm probe. It was concluded that the sensing depth of the probe was 0.28 mm at 500 MHz. This confirmed that a commercial probe can, in fact, be used for measuring the mammary tissues of the rat. Exploitation and dissemination of this work were carried out on several scientific conferences including PIERS 2018, MMS 2018, ACES 2019, PIERS 2019, ICECOM 2019.

The next step was to test the effectiveness of the Machine Learning (ML) algorithms for the classification of the collected dielectric properties. To do so, dielectric properties of discarded renal calculi (kidney stone) samples were collected in the laboratory environment with the commercial open-ended contact probe technique. The Cole-Cole parameters were fitted to the collected dielectric properties. Cole-Cole equation is then fitted to the measured dielectric properties. Since there were three renal calculi types, multiclass ML algorithms such as k-neighbors classifier (kNN), Artificial Neural Networks (ANN), and others were trained with the data. It was found that all the ML algorithms were able to classify the data with over 93% accuracy while algorithms such as kNN performed really well resulting in over 98% accuracy. This work suggested that the multiclass classification with high accuracy is possible. The dielectric property data of this work accepted for publication in IEEE TDEI Journal and the ML results are published in the Elsevier CBM Journal.

To test the developed techniques for breast cancer application, animal experiments were performed with Sprague – Dawley female albino rats. Chemically induced mammary carcinoma tissues were measured along with healthy counterparts. Collected dielectric properties were analyzed. Note that dielectric property analysis was also performed on other datasets as well including for kidney stone and blood glucose levels (published on MDPI Diagnostics journal in 2019). The collected S parameter response was then processed with the ML algorithms and it was concluded that good accuracy (over 93%) could be reached with using the S parameter response.
In principle, the tissue type can be determined based on the measured dielectric properties by using an open-ended contact/coaxial probe technique. However, the reported error rates of the technique are higher than the inherent dielectric property discrepancy between the tissues. This project revealed that with the adoption of ML algorithms, it is possible to reach over 98% accuracy during tissue identification. Furthermore, over 93% accuracy can be achieved by using only the S parameter responses. Finally, we concluded that the large majority of the information is carried in a single frequency. These results reveal that a simple Radio Frequency circuit operating at a narrow band can support the probe. Indicating a dramatic decrease both in the bulk of the device also in the cost. The next step for this project is to realize a prototype and starting the clinical trials.

A dramatic decrease both in the device size and cost will enable deployment of the device to clinics around the world including rural regions. After clinical trials, with the deployment of the device to clinics, we expect an increase in data accumulation which will support higher accuracy rates. By growing the confidence in the device through increasing the accuracy values, the biopsy cost and decision time can be decreased. This would encourage the screening of women for breast cancer more often and enable the early diagnosis of breast cancer. Moreover, with the utilization of the device for surgical margin detection, the cancer tissue can be excised with high confidence minimizing the demand for radical mastectomies.

Finally, it is known that anomalies affect the dielectric properties of the tissues. Therefore, this technique can also be applicable to other cancer/anomaly types.
presentation1.jpg