Grant Funding 2020-2022
Ministry of Education and Science of the Republic of Kazakhstan
The aim of the project: To develop an intelligent breast cancer diagnosis system. The works proposed include 1) classification and the development of a repository of thermograms, medical conditions, lifestyles, dietary, genetic information along with other cancer development symptoms; 2) development of technologies that apply machine learning and deep learning techniques to data mine the repository to predict the early development of the cancerous tumor inside the breast. 3) combination of machine learning with inverse thermal modeling to extract patient-specific tissue parameters for machine learning and find tumor sizes and locations for precision diagnosis.
Breast cancer is a multifactorial disease, the development of which is associated with changes in the genome of the cell under the influence of external causes and hormones. It is considered to be one of the most common diseases that lead to death among females. Recovery rate varies depending on the stage at which the disease is diagnosed. Breast cancer in the early stages is asymptomatic and does not cause pain. Early diagnosis is vital, as the tumor is highly treatable at the earlier stages. There are many techniques for cancer diagnosis and the most common are breast examination and mammography. it has harmful radiation, which can cause the growth of the tumor. It is therefore necessary to examine alternative diagnostic techniques to identify cancer at an early stage without harmful after effects so as to increase the chances of recovery through proper treatment. Thermography is an imaging technique using infrared rays to produce color pictures of the temperature distribution fields. This technique is suitable for the identification of tumor at the early stages as one of the tumor development signs is the increase in breast tissue temperature. Although thermography has simple working principle, the diagnosis is based on qualitative principles and human judgement, for example, asymmetry of two breasts, hyper-hermetic patterns and abnormal vascular patterns. Quantitative information is usually manually extracted by observing the temperature distribution and randomly matching the temperature profiles at different locations. With the rapid development of computer technologies, computer-aided tools can be used to support the interpretation of thermal images and help doctors to automatically identify locations and sizes of the tumors, blood perfusion and other personalized properties of the breast tissues, and thus assist with the diagnosis. In the present work we propose to combine both Bayesian Networks (BNs) and Neural Networks (NNs) with efficient inverse thermal modeling to achieve highly efficient and patient-specific precision diagnosis. Our method will rank the suggestions of both techniques and in this way we will be able to select for different medical decisions/classifications the optimal approach. More specifically, both BNs and NNs will be used for the unsupervised learning for clustering as well as for the supervised learning that will generate the prediction models for tumor. Then the efficient inverse thermal modeling will generate patient-specific tissue parameters and tumor location and size with high precision, which we are currently developing in a separate research project.
The system for early breast cancer detection and prediction will revolutionize breast cancer detection and treatment in Kazakhstan and beyond. Since thermography has not been utilized in Kazakhstan, the project will help to spawn a new thermography-related technology industry in the country. Commercialization possibilities in the global medical market include deep learning based technique for breast cancer detection. This is an area where engineering and technology has a direct role in improving the health condition of Kazakh women. A whole new thermography-related technology industry in the country can be thus promoted and generated. Work will be published in the local media and results will be presented at international conferences, published in good JCR international journals
First we complement thermography with numerical analysis in order to use patients’ personalized data such as precise breast geometry and temperature patterns to detect tumors inside the breast. To improve the accuracy and reliability of computer-aided diagnosis of breast tumors, this study used realistic 3D breast geometry based on 3D scanning, and finite element numerical modelling, which is then validated by experiments through fabricating the breast using 3D printing and molding. Thus, it can be concluded that the major important factors for precision tumor detection are tumor depth and breast geometry.
The diagnosis of the tumor inside the breast based on the thermal patterns on its surface is an “inverse modeling” problem dependent on personalized information of the patient. Inverse modeling is based on the idea that the surface thermal pattern of the breast can be used as input to determine the tumor features based on physical and physiological principles. Consequently we developed a well-validated inverse thermal modeling framework that could be used to determine the depth and size of tumor inside a breast as well as personalized patients’ breast data, such as thermograms and 3D geometry using efficient design optimization techniques such Multi-objective Genetic Algorithm (MOGA) and Finite Element Modeling (FEM).
The results show that the combination of 3D breast geometries, thermal images, and inverse thermal modeling is capable of estimating patient/breast-specific breast tissue and physiological properties such as gland and fat contents, tissue density, thermal conductivity, specific heat, and blood perfusion rate, based on a multilayer model consisting of gland and fat. Moreover, this tool is able to calculate the depth and size of the tumor, which is validated by the doctor’s diagnosis.
Furthermore we believe that machine learning technologies can potentially be used to support the interpretation of thermal images and help physicians to automatically determine the locations and sizes of tumors, blood perfusion, and other patient-specific properties of breast tissues. The main innovation of our work is the use of breast thermograms from a multicenter database without preprocessing for binary classification, assisted by the digital breast models developed and reverse modeling mentioned above. The results of the study show the usefulness and efficiency of deep learning for standardized analysis of thermograms. It is found that the model developed can have an accuracy of 80.77%, sensitivity of 44.44 % and the specificity of 100%. The limitation of the present study is that the patients’ data available for the analysis were less than the amount of data typically collected for deep learning. In addition the positive predictive value (PPV) is still considered low, which can be further improved via physics-informed Neural Network (PINN) models in the future which are being developed by us.
|Effective start/end date||1/1/20 → 12/31/22|