چكيده به لاتين
Today, doctors are able to diagnose the exact type and location of diseases by using radiography images. Among the types of medical images, we can mention MRI, CT and Ultrasound images. However, the quick diagnosis of the disease is one of the vital and decisive factors in the process of treatment and prevention of disease progression. Intelligent disease diagnosis systems using medical images are a great help in speeding up the diagnosis of the type and even the exact location of the disease.
In order to train smart models for disease diagnosis, many labeled medical images are needed so that a model can be trained using them. However, collecting and labeling a large amount of data in this area is very expensive. For this reason, using self-supervised methods to use images without labels can have a significant effect.
In this work, we have examined various self-supervised methods for training a deep neural network. These methods extract rich and useful features from unlabeled images to reach a more suitable output for disease diagnosis by using less labeled data. The task investigated in this thesis is the segmentation of medical images for brain tumor extraction, which is done on MRI images.
The deep neural network used in this thesis is Unet[1], which was introduced as a segmentation network in 2015 and has undergone many changes in the past years to improve the results. In the first part, by using auxiliary tasks such as rotation and inpainting, it has been shown that the use of self-supervised tasks can increase the Dice score by 7% to 10% compared to the time when these tasks were not used. In the second part, with inspired by contrastive learning, a new method has been introduced in the network training, which has a significant effect on learning the segmentation of medical images. There are 1000 brain volumes in the data used in this work. When we use all these 1000 volumes without self-supervise learning, the Dice score of the segmentation is 86%. Using this method, using only 300 volumes, the accuracy of the segmentation network has reached 84%.
