Neural network to improve quality of medical images
A KFU-made algorithm is already in use at the University Clinic.
The project involves Director of the Institute of Computational Mathematics and IT Dmitry Chikrin, Head of the High-Performance Computing Lab Dmitry Tumakov, Head of the Department of Medical Imaging of the University Clinic Diaz Galimyanov, Junior Research Associate Alisher Zhumaniezov, and master student Zufar Kayumov.
As Dmitry Tumakov explains, the algorithm presents a synthesis of neural networking and mathematical approaches, including Fourier transforms.
Mammography is the most effective and widely used test for the early detection of breast cancer (BC). However, early breast cancer can be very small (several millimeters) or located on a dense background, which makes it difficult to detect them during visual analysis of mammograms by a radiologist and leads to their untimely diagnosis. The question is covered in a publication by the mentioned co-authors in J. Imaging.
The algorithm consists of three stages. The first is defect detection. The second involves improving and equalizing the contrasts of various parts of the image outside the defect. The third involves restoring the defect area using a combination of interpolation and an artificial neural network. The mammogram obtained as a result of the application of the algorithm shows a significantly better image quality and does not contain distortion caused by a change in the brightness of the pixels. In total, 98 radiomic features are extracted from the original and received images, conclusions are drawn about the minimum differences in features between the original image and the image obtained by the proposed algorithm.
To increase the efficiency of mammogram analysis, a computer diagnostic system was also developed, which provides the ability to detect poorly visible changes. In a large sample (more than 600 cases), this system was shown to detect 90 percent of breast cancer cases (including 48 percent of invisible and 87 percent of poorly visible cases, on average, two years before the actual diagnosis).
“The algorithm can be improved. We managed to achieve a high quality of the image, but have there been any losses that are significant for doctors? When we talk about medical images, they are analyzed by a doctor. We also compared the results of mammography image processing obtained using a convolutional neural network model with those obtained using a nested contour algorithm model. The neural network analyzes the mammogram and finds flaws. Thus, the physician is tipped about the possibility of cancer. This is how we check whether we have lost something significant by improving the quality of the images using the developed algorithm. We managed to preserve the textures of the images, which is important for doctors,” concludes Tumakov.
The innovation has already been introduced at the University Clinic.