Projects

2022-

Development and Implementation of a Deep Learning-Based Fully Automated Approach for Automatic Detection and Measurement of Aortic Valve Calcium Score from Echocardiographic Images (TUBITAK 1001) - Ongoing project

Researcher: Dr. Ögr. Üyesi Elif Baykal Kablan

Funded Student(s): Merve Nur Çakir, Sefa Keklik

Supervisor: Prof. Dr. Murat Ekinci

Abstract
Aortic valve calcium score is widely used in the diagnosis, treatment, and follow-up of aortic stenosis and in determining the risk of coronary artery disease. Current guidelines recommend considering the aortic valve calcium score, especially in the diagnosis of low-flow and low-gradient aortic stenosis. The gold standard method for determining the aortic valve calcium score is computed tomography (CT). Agatston score is a semi-automatic method used routinely to calculate calcium score in patients undergoing cardiac CT. In Agatston scoring, the calcium area is multiplied by a coefficient determined according to the CT density. It is a method whose results vary depending on the extraction technique and is affected by minor changes in slice thickness. However, since it is the first described and the most widely used method for many years, it is still frequently preferred to benefit from the data in the existing literature. On the other hand, CT is an expensive and radiation exposing examination. Echocardiography (ECHO), which is a cheaper and radiation-free ultrasonographic method, can also be used instead of CT. There are several studies to measure the aortic calcium score with (ECHO), but all these studies are observational and semi-quantitative. In addition, the aortic valve calcium score, which is evaluated with ECHO, is a visual evaluation and does not provide an objective result, as there will be intra-observer and inter-observer diagnostic differences. Therefore, automatic measurement of the aortic valve calcium score is important.

In this project, firstly, R&D studies will be conducted to develop deep learning approaches for estimating aortic valve calcium score from ECHO images, based on calcium scores obtained from CT images by Agatston method. In this way, it is aimed to obtain the aortic valve calcium score more objectively, faster, easier, less costly, and fully automatically without exposure to radiation. The fact that the patient is currently undergoing both ECHO and CT for the diagnosis of aortic stenosis exposes the patient to extra radiation. Another study we will carry out for this purpose will be the production of artificial CT images from ECHO images. Thus, the aortic valve calcium score will be determined on the CT images to be produced without exposing the patient to extra radiation. As a result of this study, it will be investigated how close to the calcium scores obtained from real CT images can be achieved with ECHO alone or with ECHO + artificial CT.

In line with these purposes;

• A unique database including the ECHO images, tomography images and calcium scores calculated by the Agatston method by two expert radiologists from the tomography images of the patients will be created.
• Since the raw images obtained with the ECHO have redundant information on the outer edges, the regions of interest including 3 leaflets forming the aortic valve should be cropped automatically. First, regions of interest (RoI) will be manually marked by two expert cardiologists in order to provide the necessary datasets for the development of the automated system. Then, calcification formations within these regions of interest will be marked manually by expert cardiologists.
• Novel network architectures will be developed for the automatic detection of the aortic valve region in ECHO images and then the automatic segmentation of calcification formations will be developedin this region.
• It will be classified by measuring the aortic valve calcium score (Agatston score) with the original regression network architecture to be proposed according to the aortic region of interest and calcium segmentation result obtained.
• In addition, novel network architectures will be developed to generate artificial CT images from ECHO images.
• Performance measurement will be evaluated in three aspects: (i) aortic valve regions obtained with the proposed specific aortic valve detection approach from ECHO images and valve regions manually marked by expert cardiologists will be compared. (ii) aortic valve calcification regions obtained with the suggested segmentation approach from echocardiography images will be compared with the calcification regions manually marked by expert cardiologists. (iii) The score values to be calculated on the aortic region of interest and calcification region obtained will be compared with the Agatston score values that will be calculated semi-automatically by expert radiologists on CT images. (iv) The similarity between the generated artificial CT images and the real CT images will be calculated. (v) The calcium score values to be obtained from the artificial CT images produced will be compared with the score values to be obtained from the real CT images.

The originality and contributions of the project can be summarized as follows:

• To our knowledge, there is no fully automated algorithm/approach in which calcium score is measured with ECHO images in the literature.
• Also, there is no algorithm/approach to generate artificial CT images from ECHO images.
• With the algorithm to be developed, it will be possible to evaluate the aortic valve calcium score, which has a wide clinical usage, more objectively, quickly, cheaply and in a way that will not be exposed to extra radiation.
• The results of this study will shed light on the development of original, innovative, and qualified algorithms by continuing academic interdisciplinary R&D activities in this field. It may also be possible to integrate the proven artificial intelligence-based algorithms to be developed for this purpose into ECHO devices as software.

Since the project goals and studies can be a universal study in the echocardiography literature, it will provide us with important gains in national knowledge, in the direct application of interdisciplinary academic R&D studies and in the international arena.

2018-2021

Automatic Light Microscopy Scanning and Analysis based on Computer Vision and Machine Learning for Cytopathologic Differential Diagnosis of Malignant Neoplasia and Reactive Mesothelial Hyperplasias (TUBITAK 1001) - Finished project

Principal Investigator: Prof. Dr. Murat Ekinci

Funded Student(s): Elif Baykal Kablan, Hülya Dogan

Abstract
Cytomorphological screening, performed by experts under light microscope, reguires scanning all cells in the field, determining whether the cells are malignant, benign or atypical and detailed anaysis of these cells. Differential diagnosis of malignant neoplasia and reactive mesothelial hyperplasia may be difficult even for experienced cytopathologist. Also in undiagnosed cases, it is utilized from some helper time-consuming techniques such as histochemical, immunohistochemical and electron microscopy.

In this project, two and three dimesinonal (2D/3D) high resolution panoramic imaging will be obtained with scanning all field of specimen enabling optimal focusing for medical light microscope. Cell nuclei detection, cell (nuclei and stoplysm) segmentation and cell classification will be performed by developing novel computer vision and machine learning approaches in this work. Adding machine learning and compter vision approaches on human-computer interaction, success and reliability of the system about differential diagnosis of malignant and reactive cells will be increased.

Main goals of the project can be summarized as follows:

• Developing hardware and software providing optimal focusing with smart control of the movements of the microscopy platform along Z-axis to produce 2D and 3D imaging with optimal focusing.
• Medical sample is screened in XYX domain to produce an high resolution 2D/3D panaromic imaging of the specimen in different scales by making a novel studies in this project.
• Each cell (nuclei and cytoplasm) in the 2D/3D panoramic imaging obtained from the specimen screening will be segmented.
• Cell segmented will then be classified as a malignant, bening or reactive cell.
• Population distribution of cells classified as malignant cells will be analyzed for diagnosis of malignant neoplasia or reactive mesothelial hyperplasia.
• All steps in the algorithm will firstly be developed on off-line databased collected from real samples. Then real-time application of the whole system will be achieved.
• Digital 2D/3D panoramic imaging database will be collected from the samples screened by using different microscope lens. When an user is manually select a region in a panaromic image, a corresponding region to the selected area will be automatically showed by matching sub-images in the panaromic image created by different lens screening. This may be an option to creat a digital database instead of keeping the medical specimens.

The novelties in the proposed project can be listed as in below.

• Computer-aided differential cytopathological analysis of the malignant neoplasia and reactive cells is performed for pleural effusion samples.
• In the literature, focusing is supposed to be optimal but actually not. On the contrary, in this project high resolution panoramic images are obtained providing optimal focusing in the whole process.
• Specimen screening in different scales is performed by using different scales of the panoramic images.
• Instead of applying cell analysis and diagnosis on different fields of specimen, it is performed on high resolution panoramic whole image of the specimen.
• This application is applicable on different applications in microscopy.

Project objectives and studies can be a universal qualitative study in medical microscopic imaging and analysis literature, which will lead to national knowledge accumulation, transfer of academic R & D work to direct implementation, and significant gains on the international arena.

2008-2010

Avuç Izi Biyometrik Özelligine Dayali Otomatik Kisi Tanima Sistemi (TUBITAK)