Deep neural networks have repeatedly made the headlines of scientific news in recent years. The progress made in fundamental research in the last decades have now been transposed into applications in the medical domain. The automatic classification of voxels in medical images using Convolutional Neural Networks (CNNs) allows, among other things, to follow the development of an organ or the evolution of a disease.
Availability of medical imaging data for training learning algorithms have been an inherent limitation of the domain. The high cost and expertise required to annotate images to produce the segmentation ground truth are facts making datasets small, often composed of only a few images. This directly impacts the generalization capability of learning algorithms. A way to alleviate this limited availability is in image normalization accross multiple imaging domains. An imaging domain can be the site where the image is acquired, an imaging protocol or the type of machinery used to acquire the image. Conventional approaches are customarily utilized on a per-dataset basis. This strategy, however, prevents the current normalization algorithms from fully exploiting the intrinsic joint structual information and intensity scales available across multiple datasets. Consequently, ignoring such joint information has a direct impact on the performance of segmentation algorithms. The correction of the image’s intensities by an intelligent normalization process that aligns the tissues’ distributions on a common scale could allow the learning of a specific task over many different datasets. This could have for consquence to improve the data availability and the generalization capacity of learning algorithms.
This paper proposes to revisit the conventional image normalization approach by instead learning a common normalizing function across multiple datasets. Jointly normalizing multiple datasets is shown to yield consistent normalized images as well as an improved image segmentation. In our method, three networks are trained simultaneously to learn the relation between an image of any domain and a normalized output. The optimal transfer function binding the image intensities across the different datasets is learned. This function has for goal to output a precise segmentation map and realistic normalized intermediate image. The images are said realistic because the changes made by the function doesn’t affect neither the structures nor the capacity of the image to be interpreted by a clinician. These important characteristics sought in order to make the intermediate image produced understandable are made possible by using an adversarial architecture. It’s a simple, yet scalable architecture which complexity doesn’t increase when increasing the number of imaging domains.
We evaluated the performance of our normalizer on both infant and adult brain images from the iSEG, MRBrainS and ABIDE datasets. Results reveal the potential of our normalization approach for segmentation, with Dice improvements of up to 57.5% over our baseline which consists of training and testing a segmentation algorithm using different datasets. Our method can also enhance data availability by increasing the number of samples available when learning from multiple imaging domains.
Delisle, P.-L. (Author),
Lombaert (Supervisor) &
Desrosiers (Co-supervisor),
3 May 2021Student thesis: Master's thesis › Master in Engineering: Information Technology Engineering