Session Index

S6. Biophotonics and Biomedical Imaging

Biophotonics and Biomedical Imaging II
Friday, Dec. 2, 2022  15:15-17:00
Presider: Yu-Chun Lin、Sheng-Hao Tseng
Room: 2F A202
15:15 - 15:45
Manuscript ID.  0900
Paper No.  2022-FRI-S0602-I001
Invited Speaker:
Sheng-Hao Tseng
Model-driven diffuse reflectance spectroscopy for retrieving various local and systematic functional parameters of the human body
Sheng-Hao Tseng, National Cheng Kung University (Taiwan)

Diffuse spectroscopy, a variant of diverse spectroscopic methods, has been used for investigating tissue properties for decades. This technique can work with proper models to noninvasively quantify chromophore concentrations of bulk tissues. Models designed for deep tissue interrogation have been established to enable the application of diffuse spectroscopy for studying the dynamics of functional parameters of deep tissues, such as stimulation-induced hemodynamics of the brain or muscle, or the variation of water and fat concentrations of breasts caused by chemotherapy. On the other hand, due to the strong stochastic nature of light propagation in turbid media near the light source, building models for superficial tissue studies have been a challenging task and thus this topic has been vastly studied in recent years. In this talk, I will discuss the attempts we have made to develop useful models that can work in conjunction with specialized diffuse reflectance spectroscopy configurations for the effective evaluation of functional parameters of superficial tissues such as skin collagen, hemoglobin, bilirubin, and glucose. It will also be illustrated how these local and systematic functional parameters are related to the monitoring or diagnosis of numerous diseases such as keloid, psoriasis, jaundice, and diabetes.

  Preview abstract
15:45 - 16:00 Award Candidate (Paper Competition)
Manuscript ID.  0720
Paper No.  2022-FRI-S0602-O001
Pei-Chia Tsai Deep Learning for Automatic Neural Canal Opening Detection with Differentiable Spatial to Numerical Transform
Pei-Chia Tsai, National Yang Ming Chiao Tung University (Taiwan)

Our previous work has achieved automatic neural canal opening (NCO) detection by deep learning model with fully connected (FC) layer connected after a convolutional neural network (CNN). Considering the lost spatial information of the FC layer and generality of the model, we replace the FC layer with the differentiable spatial to numerical transform (DSNT) layer in this work. Our model converges faster with an accuracy of 97% and an intersection over union (IoU) of 0.85, which is close to the accuracy (97%) and IOU (0.87) of previous work.

  Preview abstract
16:00 - 16:15 Award Candidate (Paper Competition)
Manuscript ID.  0460
Paper No.  2022-FRI-S0602-O002
Yun-Jie Jhang Deep Unsupervised Learning for Image Enhancement in Nonlinear Optical Microscopy
Yun-Jie Jhang, National Tsing Hua University (Taiwan); Xin Lin, Shih-Hsuan Chia, Zi-Ping Chen, National Yang Ming Chiao Tung University (Taiwan); Wei-Chung Chen, I-Chen Wu, Kaohsiung Medical University (Taiwan); Ming-Tsang Wu, Kaohsiung Medical University (Taiwan), Kaohsiung Medical University Hospital (Taiwan); Guan-Yu Zhuo, China Medical University (Taiwan); Hung-Wen Chen, National Tsing Hua University (Taiwan)

We present an unsupervised model to enhance images in nonlinear optical microscopy. Without statistical assumptions, it can be applied to unseen samples with various hardware settings which shows significant improvement in the image-enhancement task.

  Preview abstract
16:15 - 16:30 Award Candidate (Paper Competition)
Manuscript ID.  0300
Paper No.  2022-FRI-S0602-O003
Md Azaharuddin Ansari Modification of Dynamic Lighting System to Meet Canadian Safety Standards
Md Azaharuddin Ansari, Apoorv Chaudhari, Yuan Ze University (Taiwan); Jonathon David White, Yuan Ze University (Taiwan), McMaster University (Canada); Nafia AL-Mutawaly, McMaster University (Canada)

Quality of life can be improved in people living with dementia (PWD) by reducing the sleep disturbance. Studies show that the 24-hour light cycles is a key trigger affecting sleep. We developed a dynamic lighting system to generate uniform white light using 6 different colored LEDs that varies over 254 hours. The intensity at wavelength (λ)~480 nm maximizes at noon, reduces during the evening hours and is minimal at night causing a person wakeful. This system was redesigned to meet Canadian and North American safety standards.

  Preview abstract
16:30 - 16:45 Award Candidate (Paper Competition)
Manuscript ID.  0410
Paper No.  2022-FRI-S0602-O004
Nazish Murad A Hybrid Algorithm for Imaging Breast Tumor(s) in Highly Scattered Non-Homogeneous Medium
Nazish Murad, Min-Chun Pan, National Central University (Taiwan)

The aim of this paper is to demonstrate a novel architecture of deep learning that addresses the prediction of breast cancer when it pertains to diffuse optical imaging (DOI). We were able to localize structure representations in larger areas more effectively than in smaller structures using our hybrid architecture. This study is among the foremost to specifically address signal and image domains combined in diffusion reconstruction in DOT. The network connects two encoders with one decoder path, which optimally utilizes more information from signal data and raw images.

  Preview abstract
16:45 - 17:00 Award Candidate (Paper Competition)
Manuscript ID.  0800
Paper No.  2022-FRI-S0602-O005
Ying-Ju Chen Quantitative Differential Phase Contrast Microscopy for Phase Retrieval with Self-supervised Neural Network
Ying-Ju Chen, Sunil Vyas, Hsuan-Ming Huang, Yuan Luo, National Taiwan University (Taiwan)

Quantitative differential phase contrast (QDPC) microscopy plays an essential role in the biological field since it visualizes thin transparent samples. Tikhonov regularization is typically used for phase reconstruction in QDPC imaging. However, the regularization parameter needs to be tuned manually and the selection of the parameter influences reconstructed phase images. We proposed a self-supervised deep learning-based algorithm using deep image prior (DIP) to solve this issue. The algorithm can predict phase without pre-training with a dataset and ground truth images with DIP. The proposed method was simulated, and a standard phase target was used to validate the method's feasibility.

  Preview abstract