ICB Seminar

24.06.2019

Pre and Post hoc Interpretation of Malignancy from Breast DCE-MRI

Prof. Gustavo Carneiro
School of Computer Science, The University of Adelaide
ICB Seminar room 121 at 11:30

In this presentation, we will show a systematic comparison between pre hoc (detection of lesions, followed by their classification) and post hoc (whole-image classification, followed by the visualisation  of lesions) approaches for breast screening from DCE-MRI. We aim to answer the following research questions: 1) which approach should be chosen if the goal is to optimally classify a whole scan in terms of malignant or non-malignant findings, and 2) how accurate is the localisation of malignant lesions produced by post hoc approaches when compared with the localisation of malignant lesions produced by pre hoc methods.

The pre hoc system considered in this paper is based on our recently proposed attention model [Maicas et al.,MICCAI'17] that achieves state-of-the-art (SOTA) lesion localisation, while reducing the inference time needed for traditional exhaustive search methods.  For the post hoc system, the classification is trained using a meta-learning approach [Maicas et al.,MICCAI'18] and holds the SOTA for the problems of breast screening from DCE-MRI. Decision interpretation is based on our recent 1-class saliency detector, especially designed for the weakly supervised lesion localisation problem after performing volume diagnosis.  Experiments on a breast DCE-MRI dataset containing 117 patients and 141 lesions show that the post hoc system achieves better malignancy classification accuracy than the pre hoc method. In terms of lesion localisation, the post hoc approach shows less accurate performance compared to the pre hoc system.

 

More information about the speaker.

We use cookies to improve your experience on our Website. We need cookies to continuously improve the services, to enable certain features and when embedding services or content of third parties, such as video player. By using our website, you agree to the use of cookies. We use different types of cookies. You can personalize your cookie settings here:

Show detail settings
Please find more information in our privacy statement.

There you may also change your settings later.