Deep learning applications in medical imaging - Deep Convolutional Generative adversarial networks
MetadataShow full item record
Medical imagingMedical imaging makes it possible to examine internal human tissue without performing surgery. Common medical imaging techniques include magnetic resonance (MR), computer tomography (CT), X-ray and ultrasound (US). They are all based on different technology but serve the same purpose, namely internal medical examination. Every medical imaging technique has its benefits and drawbacks; no method is perfect. In addition, some form of image processing is always necessary to visualize the images for human interpretation. A lot of effort is a put into image processing techniques to improve and enhance the image quality. Deep learning and its applicationsIn the last decade, Deep Learning (DL) has been developed and used extensively in a wide range of fields, especially in image processing tasks. DL algorithms based on convolutions in many layers, deep convolutional neural networks, have been able to tackle tasks people believed computers never would be able to handle. Some examples are face recognition, self-driving cars and automatic image diagnosis systems. One reason for the success of DL is the massive computing power of graphic processing units (GPUs) in modern computers. A computer performs many more calculations per time unit than the computers were capable of doing a couple of years ago. One exciting application of DL is image-to-image transformation. This could be transforming a landscape scene from a daylight to a nighttime setting or generating new realistic-looking faces as a mixture of other faces. Many implications are yet to be discovered and realized by this technology. The goal of this master thesis was to investigate DL applications in medical imaging. The focus was on image-to-image transformations such as noise reduction and conversion of MR and US images. Noise reductionA main issue in medical imaging is noise corruption. Removing noise from ultrasound images in real-time is therefore a very useful application. Complex noise reduction algorithms are not be able to run in real-time today. Such an algorithm was used in this thesis to generate a noise reduction reference for the DL algorithms to aim for. The DL algorithms can run on highly optimized implementations on both CPUs and GPUs. In this thesis, the trained DL algorithms were gradually modified to be able to run even faster. The result was a DL algorithm which can perform noise reduction with similar accuracy as the complex noise reduction algorithm in real-time. In addition, the DL algorithm was extended to accept a parameter to adjust the aggressiveness of the noise filtering. This was done to make the DL algorithm able to adapt to different levels of noise. MR and US image transformationMR is known for being an accurate imaging technique with high quality. However, MR is time-consuming, expensive and requires a lot of space. In this thesis MR to US image transformation and vice versa was investigated. That is to generate a realistic looking MR image from an US image. This was done by having MR and US images of the heart from the same patients. One potential application of this is multi modal image registration, which is to anatomically combine information from US and MR.