Part-time
09/03/2021
Loïc BRANSTETT
Maître de stage: Faïçal Selka
Application du Deep Learning pour l’amélioration de la qualité des images fluoroscopies à faible dose en rayon X
Introduction
Loïc BRANSTETT
20 ans
3ème année à EPITECH STRASBOURG
Auto-présentation
Introduction
Mon projet professionnel
Introduction
Choix du stage
Contexte
Environnement
Contexte
Environnement
Ma contribution
Context
Ma contribution
Context - Problème
M
Ma contribution
Context - Problème
-
Ionizing radiation and cancer risk: evidence from epidemiology. 1998 Radiat
-
Radiation Exposure to the Surgeon and Patient During a Fluoroscopic Procedure: How High Is the Exposure Dose? A Cadaveric Study. 2016 Spine
-
Radiation exposure during fluoroscopically assisted pedicle screw insertion in lumbar spine. 2000 Spine
-
Does surgeon experience influence the amount of radiation exposure during orthopedic procedures? A systematic review. 2019 orthopedic review
Ma contribution
Context - Solutions
- Breast Cancer in Women With Scoliosis Exposed to Multiple Diagnostic X Rays. J Natl Cancer Inst 1989;81:1307-1312
- Fluoroscopically Guided Interventional Procedures: A Review of Radiation Effects on Patients' Skin and Hair. Radiology 2010;254:2
- FLUOROSCOPY DURATION IN ORTHOPEDIC SURGERY. Rev Bras Ortop. 2015;46(2):136-138. Published 2015 Dec 6.
- Radiation exposure during pedicle screw placement in adolescent idiopathic scoliosis: is fluoroscopy safe?. Spine 2006
Low dose
Moins de radiation
Image bruité
Ma contribution
Context - Solutions
Ma contribution
Mon rôle
Ma contribution
Mon rôle
Low dose
Méthode classique
Deep Learning
Débruitage des images
Ma contribution
Mes tâches - Noise2Noise
PSNR: 30.5 db
- Noise2Noise: Learning Image Restoration without Clean Data. 2018 ICML
Noise2Noise
Ma contribution
Mes tâches - Base de donnée
x2000
x6000
Speckle
Gaussian
Poisson
Add noise
Input
Ma contribution
Mes tâches - Noise2Void
- Noise2Void - Learning Denoising from Single Noisy Images, 2019 arXiv
Ma contribution
Mes tâches - Intégration C++ libTorch
$ ./n2n model.pt input.png output.png
Cuda: avalaible
Module: 1348ms
Input: 34ms
Prediction: 36ms
Output: 29ms
Bilan du stage
Résulats
PSNR: 35.57 db
PSNR: 13.21 db
GT
Bruité
N2N
Bilan du stage
Compétences
Conclusion
FIN
Bilan du stage
Bilan du stage
