Published on February 24, 2023–Updated on March 2, 2023
Dates
on the March 3, 2023
from 10:30am to 12:00pm
Location
Inria Sophia Antipolis
Program
10:30 - 11:00
Thi Khuyen Le (Chair of O. Humbert)
Université Côte d'Azur
Comparison of handcrafted radiomics and 3D-CNN models to diagnose striatal dopamine deficiency in Parkinsonian syndromes based on 18F-FDOPA PET images
18F-FDOPA PET imaging is highly sensitive for detecting striatal presynaptic dopaminergic denervation and is used in clinical practice to assist the diagnosis of degenerative parkinsonian syndromes. Using 18F-FDOPA PET images, we investigated the respective performances of semi-automatic handcrafted radiomic and automatic 3D-CNN models to detect striatal dopaminergic denervation. We
demonstrated that the performance of the 3D-CNN model is high and stable (balanced accuracy = 99.18%) on an internal test set of PET images (n = 100 patients) acquired with a different PET scanner and a different imaging protocol from the training/validation set (n=417 patients). The 3D-CNN model correctly classifies PET images without the segmentation step required for the handcrafted radiomic model. It led to diagnostic performances similar to those of a junior medical expert. Furthermore, in a non-expert center, the 3D-CNN model allowed physicians to correctly reclassify 10 patients out of a cohort of 170 patients (6%), without introducing additional diagnostic errors.
10:30 - 11:00
Irene Balelli
External researcher
Inria, Epione team
The search for causality in the analysis and modeling of biomedical data
Current evolution in digital health heavily relies on data-driven machine learning and deep learning methods, which despite their impressive results, present some drawbacks related to 1) the need of a huge amount of curated data for training and 2) the lack of an underlined "causal" structure preventing the system to go beyond association. In this talk I will discuss the need of questioning the “How?” and the “Why?” when building data assimilation methods and predictive models, especially in the medical context, for more interpretable, explainable and trustable algorithmic prescriptions. I will present some helpful ingredients for this query and discuss some of my works, embedded in this philosophy.
