XClose

UCL Centre for Medical Image Computing

Home
Menu

Maira Tariq - Stephen Morrell

07 March 2018, 1:00 pm–2:00 pm

Event Information

Open to

All

Location

UCL Bloomsbury - Roberts 106 Roberts building

Maira Tariq

Title: Pushing the standards for cancer diagnosis with advance diffusion MRI techniques: application in prostate and rectal cancer

Abstract:

In my work, I develop biophysical models of diffusion MRI for better characterisation of cancer microstructure, using the VERDICT MRI framework. The first part of my talk describes the construction of biophysical models using ultra-high-field diffusion MRI to investigate microstructure in ex-vivo samples of prostate tissue. These images have spatial resolution approaching the cellular scale, allowing us to explore the precise microstructure basis of the various models. The second part of the talk shows preliminary results for the first clinical application of VERDICT MRI to rectal cancer. The VERDICT framework has potential to improve the currently used imaging techniques via better staging, prognostication, and evaluation of treatment response.

Stephen Morrell

Title: Large-scale mammography CAD with deformable convolutional nets

Abstract:

State of the art deep learning methods for image processing are evolving into increasingly complex meta-architectures that incorpo- rate a growing number of modules. Among them, deformable convolu- tional nets (DCN) and region-based fully convolutional networks (R- FCN) have the potential to revolutiose mammography CAD: R-FCN excels at object detection, while deformable convolution and pooling can model a wide range of mammographic findings of different shapes and scales, thanks to their versatility. In this study, we present an neural net architecture that adapts R-FCN / DCN to the mammographic image analysis problem. We motivate changes in architecture and parameters in order to bridge the gap between natural images and mammograms. We trained the network on a large, recently released dataset of mammo- grams (Optimam) including 6,500 cancerous images. By combining our modern architecture with such a rich dataset, we achieved an area under the curve of 87.9% for breast-wise detection in the DREAMS challenge (130,000 withheld images), which is higher than any other submission in the competitive phase of the challenge.