CARVAJAL, M. Juan (2023) 3D image segmentation with complex amplitude of microscopic images. Multichannel approach to U-Net architecture PRE - Research Project, ENSTA.

[img]
Preview
PDF
923Kb
[img]
Preview
Image (JPEG)
128Kb

Abstract

In order to apply all the information coming from a microscopic image, the complex field is used as input to perform medical image segmentation. This paper presents research related to the training of a 3D U-Net convolutional network, which is adapted to process multiple inputs, taking advantage of multichannel convolution. By concatenating the monochromatic variables of intensity and phase of the complex amplitude of an object, we obtain a multichannel image capable of being processed by a U-Net model as a color image, taking all the information concerning the object in the training (RI, Thickness). We performed training tests with different combinations of the described inputs and their derivatives (such as sine and cosine of the phase). Under the data simulation conditions described in the paper, we show that no significant improvement in the performance of the network can be evidenced when compared to a traditional input of the real amplitude.

Item Type:Thesis (PRE - Research Project)
Subjects:Information and Communication Sciences and Technologies
ID Code:9523
Deposited By:Juan CARVAJAL
Deposited On:23 août 2023 15:02
Dernière modification:23 août 2023 15:02

Repository Staff Only: item control page