ALT'22 B-P-9
BIOMEDICAL PHOTONICS
Enhancing diffuse optical tomography using deep learning
M.A. Ansari*, A. Meisamy Optical Bio-imaging Lab, Laser and Plasma Research Institute, Shahid Beheshti University, Tehran, Iran
m_ansari@sbu.ac.ir
Diffuse Optical Tomography (DOT) is a non-invasive imaging technique using near-infrared electromagnetic waves to measure the optical properties of biological tissue from boundary measurement. Image reconstruction in this method is an inverse, ill-posed, and nonlinear problem [1]. Traditional optimization methods can't overcome explicitly this problem. Recently, deep neural networks were used in image reconstruction and they have achieved significant improvement [2]. In this research, we used neural network algorithms to reconstruct the absorption coefficient distribution of 3-dimensional phantoms. We demonstrate that deep learning algorithms has a good performance in reconstructing DOT images in comparison to the model-based method.We generate 17000 digital cubic phantoms include inclusions with different size, shape, different places and absorption coefficients. The background absorption and scattering coefficient of the phantoms were assumed 0.01 and 1 mm-1, respectively. An imaging system including 25 sources and detectors with Gaussian profile were considered up and down of tissue. Absorption coefficient of inclusions varied from 0.02 to 0.08 mm-1 with 0.02 steps and scattering coefficients were the same with background. For solving continuous wave (CW) forward problem of the diffuse equation, we used Toast++ open-source software based on a finite-element solver [3]. We propose two different neural network architectures: a fully connected layer with 8 hidden layers and a convolutional network with 6 convolutional and pooling layers. The learning rate and epoch were set to 0.0001 and 1000, respectively. For preventing overfitting, we used the earlystopping method. Mean squared error was used as a loss function, ReLU as an activation function, and Adam as an optimizer. The models were implemented in Keras and Input data was separated into 12240 training, 3400 testing, and 1360 validation dataset. The performance of networks was evaluated by four metrics including mean absolute error (MAE), mean squared error (MSE), Peak Signal to Noise Ratio (PSNR), and Structural Similarity Index Metric (SSIM). For comparison with model-based DOT reconstruction methods, we used the conjugate gradient algorithm with Total variation (TV) regularization. Fig.1 shows ground truth and reconstructed distribution of absorption coefficient in z=25. Result shows, by using fully connected layer and convolutional neural network, MAE 76% and 69% and MSE 84% and 62% respectively were reduced and PSNR was doubled in comparison with CG. Accordingly, both neural networks have better performance in DOT image reconstruction than model-based method.
é
(a)
(b)
6
(c)
(d)
Fig. 1. a) true image in z=25, b) reconstructed image by Conjugate gradient algorithm, c) reconstructed image by fully connected neural network, d) reconstructed image by convolutional neural network.
This work is based upon researchfunded by Iran national Science Foundation (INSF) under project number 98029460.
[1] Hoshi. Y. and Yamada. Y., Overview of diffuse optical tomography and its clinical applications, Journal of biomedical optics, vol. 21, pp.091312, (2016).
[2] Jalalimanesh. M.H. and Ansari. M.A., Deep learning based image reconstruction for sparse-view diffuse optical tomography, Waves in Random and Complex Media, pp. 1-17, (2021).
[3] Schweiger. M. and Arridge. S. R, The Toast++ software suite for forward and inverse modeling in optical tomography, Journal of biomedical optics, vol. 19, pp. 040801, (2014).