Please enable JavaScript.
Coggle requires JavaScript to display documents.
Unsupervised FLIM denoising - Coggle Diagram
Unsupervised FLIM denoising
introduction
Introduce FLIM, the importance of FLIM, applications of FLIM
Challenges associated with acquisition speed. Current limitations.
Introduce denoising methods. Emphasize DL-based methods and unsupervised methods.
Introduce the proposed method. DL-based with superior performance, unsupervised (generalizability, work with limited data), light weight and relatively less computationally expensive. Flexible and opensource, easy-to-use
Give a vision of the potentials of the proposed method
Results
10 levels TMA denoise matrix with limited data
302 TMA cores with chi-square metric
Timelapse of cells at 2 different SNR levels
Eventful timelapse
Generalizability test, direct use of pretrained denoise model on testing data
Methods
blind-spot neural networks for unsupervised denoising
Stochastic blind-spot neural network
random dropout in 3D, sampled from Bernoulli distribution
data augmentation as well, prevent overfitting, work with extremely limited data
neural network design
3D convolution and 1D convolution hybrid design
1D fully-connected layer covers the whole lifetime dimension
3D convolution that consider neighbor information in both spatial and lifetime domain
Training and inference specifics
Random transformation + random dropout in 3D, prediction pixel values at blind-spots
At inference, stochastic forward pass and averaging the results
FLIM imaging platform
Sample set and dataset
discussion