Knee joint segmentation based on deep learning and label fusion techniques

• Κύριος συγγραφέας: Κελλάρη, Ασπασία
• Άλλοι συγγραφείς: Kellari, Aspasia
• Γλώσσα: English
• Έκδοση: 2020
• Θέματα: Image segmentation; Convolutional neural networks; Label fusion; Image registration; MRI; Τμηματοποίηση εικόνας; Συνελικτικά νευρωνικά δίκτυα

Image segmentation, the process of delineating objects and other regions of interest within an image, plays an important role in the medical field. The ability to accurately partition an anatomical structure into different regions is essential in numerous biomedical applications and can aid the work of medical experts in various tasks, such asstudying ana tomical structures, localizing different pathologies, quantifying tissue volumes, tracking the progress of a disease or even performing computer-aided surgeries. The problem of medical image segmentation is well studied and various methods have already been proposed in literature that offer good solutions in a variety of problems. In recent years, the focus has shifted towards deep learning methods, and especially deep networks that have exhibited a state-of-the art performance in various applications. Such networks have the ability to automatically extract meaningful patterns from the raw data, without relying on hand-crafted features and at the same time require little preprocessing compared to other methods. However deep learning methods are faced with certain limitations that arise from the medical domain. One of them is the limited number of data that can significantly compromise the performance of a deep network, that typically requires a big amount of training data to infer accurate solutions. Other problems include the high variability that exists within datasets, the lack of consistent patterns within the anatomical structures and the high dimensionality of the various image modalities. In the present text we tackle the problem of the automatic segmentation of MRI images of the knee complex, by employing 3D Convolutional Neural Networks (CNNs), that utilize the full volumetric information of the data. To address the aforementioned limitations (limited training data, high variability) that exist in the datasets, we incorporate deformable registration and a label fusion technique along with the segmentation network. Multiple networks are trained on sub-spaces, that were created by registering the training data into different reference spaces. The outputs of the segmentation networks are then combined to produce a final segmentation for each test image. The proposed method is based on the assumption that the combination of similar versions of the data across multiple targets can produce more accurate solutions.