Deep learning-based computed tomography image processing

Computed Tomography (CT) is one of the most common modalities of medical imaging. CT scan has become a useful screening tool for body composition analysis and spinal conditions. Body composition is an emerging biomarker for cancer diagnosis and treatment. The assessment of the body composition profile of cancer patients during treatment and survivorship revealed the critical role of skeletal muscle mass in drug toxicity, hospital stay, infection rate and survival outcome. In this thesis, we propose a set of automatic CT image analysis frameworks for skeletal muscle mass and adipose tissue segmentation. This pipeline includes middle cross-sectional slice classification at the third lumbar vertebra in a CT-scan volume and muscle and adipose tissue segmentation at the third and fourth lumbar vertebrae levels. A multi-class segmentation network is trained to generate the segmentation map of skeletal muscle, subcutaneous adipose tissue, visceral adipose tissue and intramuscular adipose tissue in the third lumbar vertebrae level CT images. The model is designed to translate the spatial resolution from the feature maps of the encoder section to the decoder layers and learn the data representations at various receptive fields. To develop a computer-aided detection technique for vertebral column metastases and other spine related diseases, an automatic vertebral column segmentation and identification method is essential. In the second part of this thesis, a deep-learning based method for accurate pixel-level labelling of vertebrae on CT images is proposed. This algorithm includes 3D vertebral column segmentation and localization on CT-scan volumes. A two step semantic segmentation model utilizing a pixel-link map is introduced to tackle the vertebrae identification task. The proposed methods leverage deep learning algorithms to learn the representation of the data and map the input images to the desired output. Several datasets of CT images from various clinical institutions were obtained to train and evaluate the proposed models. These methods, accelerate the CT image analysis process and provide the information required for physicians to make a diagnosis.