Efficient methods for the semi-supervised and weakly-supervised analysis of medical images

点击翻译按钮获取中文摘要

Context Recently, artificial intelligence (AI) techniques have led to substantial improvements in performance for various problems of medical imaging, including image segmentation and reconstruction. However, these techniques often need large amounts of annotated data which are rarely available in real-life applications. This scarcity of annotated data impedes their ability of generalizing to new data, thereby limiting their adoption in clinical practice. In many medical applications, large amounts of unlabeled data are often obtainable, which could be exploited in a semi-supervised setting. The limited efficiency and scalability of current approaches is, however, a major obstacle to using this mass of unlabeled data.Research objectives and methodology This research aims at developing efficient methods for the segmentation and reconstruction of medical images, that can achieve state-of-the-art performance when annotated data are limited. Toward this goal, two complimentary axes of research will be explored. The first axis proposes to investigate novel approaches based on convolutional neural networks (CNNs) that exploit unlabeled data and low-cost annotations to improve performance when there are few labeled examples. This will be achieved by incorporating strong shape priors and constraints into the training process. The second research axis will investigate efficient and scalable methods based on distributed optimization to segment and reconstruct large images. These methods will accelerate processing times by decomposing complex optimization problems into smaller and easier sub-problems that can be solved in a distributed manner. The broader vision of this program is to make AI techniques more usable in clinical practice, by better exploiting available data and leveraging powerful optimization techniques.Contributions & impact This research proposes flexible and efficient methods to transfer problem-specific knowledge into data-driven methods like CNNs. By reducing the need for annotated data, these methods will alleviate the work of radiologists and other experts, saving both time and money. Improving the segmentation and reconstruction accuracy will also give clinicians a more reliable and complete information for diagnosis and treatment. Finally, increasing the scalability of current approaches will benefit to medical imaging and various other fields, where the volume and resolution of data is increasing each year.