Fernando Pérez García - Prabhjot Kaur - CMIC/WEISS joint seminar series

Speaker: Fernando Pérez García
Title: TorchIO, a library for preprocessing and augmentation of medical images in deep learning

Abstract: TorchIO is an open-source Python library developed by researchers at UCL and King's College London to enable efficient loading, preprocessing, augmentation and patch-based sampling of 3D medical images in the context of deep learning. It follows the style of PyTorch and integrates standard medical image processing libraries to efficiently process images during training of neural networks. TorchIO transforms can be composed, reproduced, traced, inverted and extended. Multiple generic preprocessing and augmentation operations as well as simulation of MRI-specific artifacts are available. TorchIO was developed to help researchers standardise medical image processing pipelines and allow them to focus on the deep learning experiments. It encourages open science, as it supports reproducibility and is version controlled so that the software can be cited precisely. Due to its modularity, the library is compatible with other frameworks for deep learning with medical images. 

Speaker: Prabhjot Kaur
Title: Unsupervised approaches to improve quality of  MRI images

Abstract: High quality MRI images are generally desired for precise medical diagnosis and analysis. The quality of MRI images is characterized by spatial resolution, signal to noise ratio (SNR), and image contrast. Each of the resolution, SNR and contrast of MRI images is limited due to various financial and hardware based constraints. For instance, MRI images with better contrast and visual image details can be obtained when acquired from MRI scanners with high magnetic field strength (FS) as compared to low FS MRI scanners. However, the accessibility to high FS MRI scanners is limited due to its high cost.

In this talk, I will discuss the unsupervised approaches addressed for two problems: (i) improve resolution of MRI images, and (ii) the estimation of high quality MR images from low FS MR images. For the resolution improvement, the super resolution approach in sparse representation framework is addressed which does not require any example image. A novel gradient profile based constraint is proposed to reduce the smearing of image details while upsampling the MR images.

The estimation of high quality MR images from low FS MRI scanners is formulated as an inverse problem. Since the mapping relating the low FS MRI image and high quality image is not known, it is also estimated along with the high quality image using the alternate minimization framework. There can be several combinations of the mapping and  high quality image for a given low FS image. Hence, a novel constraint is developed which exploits the physics of MRI acquisition to regularize the estimation of high quality images.