3IA PhD/Postdoc Seminar #24

Landmark detection via convolutional neural network

Prostate cancer is the fourth most common cancer diagnosis in worldwide, with around 1.4 million new cases and 370 000 deaths in 2020. If prostate cancer is suspected, the following procedures may be used to decide if more diagnostic tests are needed: prostate-specific antigen (PSA) test, digital rectal exam (DRE), etc. Then further tests will be used to confirm whether a person has prostate cancer on condition that the PSA or DRE test results are abnormal. Many tests can suggest that cancer is present, but only a biopsy can make a definite diagnosis. The recent introduction of multiparametric magnetic resonance imaging (MP-MRI) now allows for imaging-based identification of prostate cancer, which may improve diagnostic accuracy for higher-risk tumors. Targeted MR/ultrasound (US) fusion biopsy is a breakthrough technology made possible by overlaying ultrasound images of the prostate with MRI sequences for visualization and targeting lesions. Suspect areas detected by the MRI are thereby displayed on the ultrasound scanner, allowing the urologist to target the necessary biopsies in real-time. Detecting 3D anatomical landmarks on both image modalities is a crucial step for targeted MR/US fusion biopsy. However, manually landmarking in 3D MR/US images is tedious, time-consuming, and lacks reproducibility. Therefore, a fast, accurate, and automatic 3D landmark detection system is meaningful for clinical applications. In this presentation, we will introduce various methods for landmark detection, and show some primary results for MR/US prostate anatomical landmark detection based on a database from cooperating company.

Detection of neural synchronization and implication for neuroscience experimental design

Two neurons are said to be synchronized when their spike trains coincide more than when they are independent. It is commonly accepted that this phenomenon plays a very important role in the neural activity. The construction of statistical tests for its detection has been the subject of much interest in the literature and in particular with the work of Albert et al. (2015) on asymptotic tests of Bootstrap and permutation. This presentation is in the same vein, and will focus on the construction of a criterion ensuring the detection of synchonization in the case of a non-asymptotic test. This criterion is constructed in such a way as to ensure control of the first and second kind errors whatever the size of the sample considered. We also apply this criterion to some classical models of interacting neurons, typically the well known jittering Poisson and Hawkes models, and deduce informations about the choice of some experimental quantities.

Joint work with: Éva Löcherbach, Patricia Reynaud-Bouret and Étienne Tanré

Open discussion on the two contributions