Speaker: Alexandre Ayotte, M. Sc. student in Computer Science, Department of Computer Science, Faculty of Science, Université de Sherbrooke
Abstract: For patients with a kidney lesion, knowledge of malignancy as well as the subtype and grade (in case the tumor is malignant), are essential elements for prognostic. Currently, the standard procedure to obtain this information is to perform a biopsy of the lesion. Biopsy is very invasive and unreliable for heterogeneous tumors, non-invasive imaging analysis such as magnetic resonance imaging (MRI) using artificial intelligence is an alternative that has attracted a lot of interest in recent years. Although recent studies show that convolution neural networks show great promise when it comes to classifying renal tumors with respect to malignancy, subtype or grade, the results are not yet convincing enough to be an alternative to biopsy.
In this project, we tested whether it is possible, using multitask learning, to improve the performance of a deep learning model for the classification of malignancy, subtype, and grade of renal tumors. We used a dataset consisting of contrast-enhanced T1- and T2-weighted magnetic resonance images of 1082 patients with a lesion in one of the kidneys, the segmentation of that lesion for each modality and a clinical dataset. We demonstrated the relevance of a single-task learning model using imaging compared to a model based on clinical data. Subsequently, we compared convolutional neural networks trained on the tasks individually to models trained to predict all 3 features simultaneously. Finally, we tested whether it is beneficial to force a model to simultaneously classify the tumor and predict a set of radiomic features that will act as auxiliary tasks.
Jury member, president : Maxime Descoteaux, Professor, Department of Computer Science, Faculty of Science, Université de Sherbrooke
Jury member, external evaluator : Julien Cohen-Adad, Professor, École Polytechnique de Montréal
Jury member, research director: Martin Vallières, Professor, Department of Computer Science, Faculty of Science, Université de Sherbrooke
All interested persons are cordially invited.