چكيده به لاتين
Ultrasound is one of the most widely used methods in internal organs and the
human body soft tissues imaging due to its characteristics such as noninvasiveness, portability, and relatively low cost. However, the low quality and
noise in ultrasound images make the need for their preprocessing for medical
diagnoses inevitable. One of the noise reduction methods in ultrasound images is
the noise reduction method based on deep learning. Convolutional neural
networks (CNNs) are one of the most successful deep learning models explicitly
used to apply to visual data .
In this study, first, using software available in medical ultrasound, a dataset
consisting of anatomical images (label data) was collected, and their ultrasound
equivalent (input data) was simulated using the field II toolbox. Then by applying
preprocessing on the images and using the U-Net algorithm, five networks with
different cost functions were trained to remove the noise of these images. 80% of
the total data were used for training, 10% for validation, and 10% for test.
Data augmentation techniques were applied in the first three networks, and a
combination of L1 and multi-scale structural similarity index (MS-SSIM) cost
functions, L1 cost function, and MS-SSIM cost function were used respectively
to compare the performance of networks. Also, the other two networks were
trained without data augmentation and using L1 and MS-SSIM cost functions to
investigate the effect of data augmentation methods in preventing overfitting in
network training. According to the values of noise cancellation metrics, the first
network with data addition and combined cost function based on Peak Signal to
Noise (PSNR) of 16.56, Structural similarity index (SSIM) of 0.51, MS-SSIM of
0.72, and feature similarity index (FSIM) of 0.76 performed the best noise
removal performance on the test data compared to other networks. In addition,
three segmentation metrics, including Dice coefficient, mutual information, and
Hausdorff distance, were used to evaluate networks' ability in this field. These
metrics showed that the trained networks perform the segmentation operation
relatively but are not accurate enough to use in precise applications such as
medical image segmentation. Finally, the selected network showed superior
performance in noise removal with PSNR value of 24.67 and SSIM value of 0.79
compared to classic denoising methods.
