Coordinate-Unet 3D for segmentation of lung parenchyma
LE, Van-Linh
Institut de Mathématiques de Bordeaux [IMB]
BoRdeaux Institute in onCology [Inserm U1312 - BRIC]
Institut Bergonié [Bordeaux]
Modélisation Mathématique pour l'Oncologie [MONC]
Institut de Mathématiques de Bordeaux [IMB]
BoRdeaux Institute in onCology [Inserm U1312 - BRIC]
Institut Bergonié [Bordeaux]
Modélisation Mathématique pour l'Oncologie [MONC]
SAUT, Olivier
Institut de Mathématiques de Bordeaux [IMB]
Modélisation Mathématique pour l'Oncologie [MONC]
Institut de Mathématiques de Bordeaux [IMB]
Modélisation Mathématique pour l'Oncologie [MONC]
LE, Van-Linh
Institut de Mathématiques de Bordeaux [IMB]
BoRdeaux Institute in onCology [Inserm U1312 - BRIC]
Institut Bergonié [Bordeaux]
Modélisation Mathématique pour l'Oncologie [MONC]
Institut de Mathématiques de Bordeaux [IMB]
BoRdeaux Institute in onCology [Inserm U1312 - BRIC]
Institut Bergonié [Bordeaux]
Modélisation Mathématique pour l'Oncologie [MONC]
SAUT, Olivier
Institut de Mathématiques de Bordeaux [IMB]
Modélisation Mathématique pour l'Oncologie [MONC]
< Reduce
Institut de Mathématiques de Bordeaux [IMB]
Modélisation Mathématique pour l'Oncologie [MONC]
Language
en
Communication dans un congrès
This item was published in
WSCG 2023 – 31th International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision, 2023-05-15, Pilsen. p. 36-42
English Abstract
Lung segmentation is an initial step to provide accurate lung parenchyma in many studies on lung diseases based on analyzing the Computed Tomography (CT) scan, especially in Non-Small Cell Lung Cancer (NSCLC) detection. ...Read more >
Lung segmentation is an initial step to provide accurate lung parenchyma in many studies on lung diseases based on analyzing the Computed Tomography (CT) scan, especially in Non-Small Cell Lung Cancer (NSCLC) detection. In this work, Coordinate-UNet 3D, a model inspired by UNet, is proposed to improve the accuracy of lung segmentation in the CT scan. Like UNet, the proposed model consists of a contracting/encoder path to extract the high-level information and an expansive/decoder path to recover the features to provide the segmentation. However, we have considered modifying the structure inside each level of the model and using the Coordinate Convolutional layer as the final layer to provide the segmentation. This network was trained end-to-end from a small set of CT scans of NSCLC patients. The experimental results show the proposed network can provide a highly accurate segmentation for the validation set with a Dice Coefficient index of 0.991, an F1 score of 0.976, and a Jaccard index (IOU) of 0.9535.Read less <
English Keywords
Lung segmentation
NSCLC
Unet
Coordinate Convolutional
Deep Learning
Origin
Hal imported