Innovative method for diagnosing thyroid nodules based on ultrasound and machine learning

Keywords: thyroid, tumors, ultrasound, neural networks, quantitative ultrasonography, classification

Project Summary

The aim of the project is to develop algorithms for processing ultrasound images to aid in the diagnosis of thyroid nodules. The motivation for the project is the fact that in many cases, diagnosis is currently imperfect, which results in a large number of unnecessary biopsies and even thyroidectomy surgeries. Such surgeries are a great burden for the patient and have serious consequences, as they require lifelong administration of appropriate medications. In addition, in some cases, patients develop resistance to these medications, which creates a serious health problem for them. While surgical interventions are justified in the case of malignant tumors, they are definitely harmful in the case of benign tumors. Increasing the effectiveness of diagnosis would reduce the number of cases in which surgery was performed unnecessarily. This would be beneficial for many reasons. First of all, patients for whom surgery is not necessary would not undergo it, and therefore would not be unnecessarily exposed to stress, pain and possible later complications. Fewer unnecessary surgeries would translate into their greater availability for people for whom surgery is necessary. Better diagnostics would also have a positive impact on healthcare costs – both surgeries and possible later complications are associated with high costs.

As part of the project, it is planned to use machine learning methods to analyze ultrasound images in order to create models that extend diagnostic capabilities and support physicians in assessing the tumor. In particular, it is planned to develop methods for segmentation (i.e. marking the area containing the tumor on the image) and for tumor classification. It is planned to use classical methods, modern deep learning methods, as well as methods combining both approaches. Also considered will be parameters related to the texture of the ultrasound image described by the GLCM matrix, as well as parameters related to the morphology of the tumor - for example, its size, shape, contrast to other tissues, edge structure, calcifications, etc.

The models will be created using ultrasound data collected in close cooperation with doctors from the Maria Skłodowska-Curie National Institute of Oncology. The research concept assumes working on two types of data. The first are standard ultrasound images – so-called B-mode images. The second are sequences of images showing changes in the thyroid gland occurring after contrast administration. The second type of data allows for determining a number of parameters related to the blood supply to the tumor, which cannot be determined from static B-mode images.