Titre :	Automatic Detection of Anatomical Landmarks for Image Registration in Radiotherapy
Type de document :	texte imprimé
Auteurs :	M.Oussama Mebarki, Auteur ; Youcef Attoui, Auteur ; Hacene Azizi, Directeur de thèse
Année de publication :	2022
Importance :	1 vol (74 f.)
Format :	29 cm
Langues :	Anglais (eng)
Catégories :	Physique
Mots-clés :	Point matching Image registration
Index. décimale :	530-Physique
Résumé :	Anatomical landmark correspondences in Radiation Therapy imaging provides extra guidance information for the medical imaging registration. However, manual landmark identification is intensive and time consuming. Therefore, developing an end-to-end deep learning approach to automatically detect landmark correspondences in pairs of two-dimensional (2D) images of Cone-Beam CT (CBCT) and planning-CT is extremely important. Our method consists of a U-net-based convolutional neural network that has been trained to recognize salient spots in both image modalities and anticipate matching probabilities for landmark pairings. We trained our method using 58 scans of 2D-axial in the pelvic area, ranging from 140 to 160 slices CTs and 86 to 88 slices in-room CTs. The training results showed promising results in terms of landmark predictions, with high accuracy and low loss scores. On the other hand, we compared the SIFT descriptor to another SIFT variant approach for matching key-points. We employed the Mean-Squared-Error (MSE) metric to compare our model to a traditional type of image registration called Elastix. Both techniques produced acceptable accuracy values, although our method is slightly more accurate.
Côte titre :	MAPH/0556
En ligne :	https://drive.google.com/file/d/15uU7N68-qxuq-6BRE20ADiyaIE9gX41G/view?usp=shari [...]
Format de la ressource électronique :	pdf

Automatic Detection of Anatomical Landmarks for Image Registration in Radiotherapy [texte imprimé] / M.Oussama Mebarki, Auteur ; Youcef Attoui, Auteur ; Hacene Azizi, Directeur de thèse . - 2022 . - 1 vol (74 f.) ; 29 cm.
Langues : Anglais (eng)

Catégories :	Physique
Mots-clés :	Point matching Image registration
Index. décimale :	530-Physique
Résumé :	Anatomical landmark correspondences in Radiation Therapy imaging provides extra guidance information for the medical imaging registration. However, manual landmark identification is intensive and time consuming. Therefore, developing an end-to-end deep learning approach to automatically detect landmark correspondences in pairs of two-dimensional (2D) images of Cone-Beam CT (CBCT) and planning-CT is extremely important. Our method consists of a U-net-based convolutional neural network that has been trained to recognize salient spots in both image modalities and anticipate matching probabilities for landmark pairings. We trained our method using 58 scans of 2D-axial in the pelvic area, ranging from 140 to 160 slices CTs and 86 to 88 slices in-room CTs. The training results showed promising results in terms of landmark predictions, with high accuracy and low loss scores. On the other hand, we compared the SIFT descriptor to another SIFT variant approach for matching key-points. We employed the Mean-Squared-Error (MSE) metric to compare our model to a traditional type of image registration called Elastix. Both techniques produced acceptable accuracy values, although our method is slightly more accurate.
Côte titre :	MAPH/0556
En ligne :	https://drive.google.com/file/d/15uU7N68-qxuq-6BRE20ADiyaIE9gX41G/view?usp=shari [...]
Format de la ressource électronique :	pdf