Medical image representations in cancer segmentation

Abstract

Purpose Cancer is the second leading cause of death globally and was responsible for an estimated 9.6 million deaths in 2018 [1]. One of the treatments used to cure cancer is radiotherapy, where a precise delineation of the cancer is crucial. The radiation therapy needs to be precise not to damage any surrounding tissues or, in the worst-case organs. However, the delineation process is both time-consuming and affected by inter-observer variability. Automating this process will free up time for the radiologist, fasten the treatment time, and remove the inter-observer variability. Theory and Method Before the machine learning model can assist with medical delineations, it must perform well. In this thesis, use of U-Net architecture is used to train the model on CT (Computed Tomography), and PET (Positron Emission Tomography) scans gathered from the University Hospital in Oslo. The dataset consists of 197 patients and is divided into three sets; 142 in the training set, 40 in the test, and 15 in the validation set. New aspects of this research are implementing two types of image representations, histogram of oriented gradients (HOG) and local binary pattern (LBP). The U-Net is run on different models using combinations of the already existing medical images and the new image representations. Several model parameters have also been tested, such as augmentation, windowing, and the use of dilated convolutions. The HOG images are created by using local contrast normalized blocks of histograms of oriented gradients. The different image representations take multiple parameters into account when created. This thesis has focused on the descriptor and normalization blocks’ different sizes when creating the HOG images. In contrast to earlier work, this thesis uses the image representation of HOG and not its descriptor vector, which is commonly used in earlier studies. LBP uses local structural changes to represent each pixel. Different radius sizes and number of descriptive points have been tested to find the optimal LBP - parameters. Results The result found that augmentation is vital for a precise delineation, increasing all the different model’s performance. The results also show that the image representations perform best when the descriptor blocks and radius are small. HOG is the best performing of the two image representations. However, they do not beat the medical images alone (0.666). The highest performance was reached when combining the medical images and HOG image representation using both the CT and PET channels (0.675). Conclusion The thesis has shown that adding new image representation to the model can improve performance. However, the score has only increased the performance by one percent, and there is a significant computational cost creating these image representations. There are still multiple ways to improve the image representations, and HOGs full potential has not been exploited. This thesis, therefore, leaves room for numerous future works regarding the usage of the HOGs image representation and possibilities for implementing other image representations for segmentation tasks.