Brain_2.5DUNet

[yingmuzhi]

Hello, nice to meet you. The experiments you made were really great and I want to find the axon tracks in human brain just like your elife paper said. I chose the GW24 dataset you reconstructed using QLIPP and two-step algorithm. After background corrected, I chose 1250 pieces of images and saprated these 2048 * 2048 pixel pictures into 256 * 256 pixel without resize, and I also made z-score standardization and Otsu thresholding(Rosin is also attempted in another experiment). Then I trained images using 2DUnet and it came out with the PCC during training about 0.80+. However in inference stage, the result was not good, and I can not find the axon track just as clear as fluorescence image. Could you help me find how to solve my problem? Is that because 256 * 256 pixel images are two big or two small, or the deep learning model I chose are easily overfitting or just can't fit this complex nonlinear transformation from label-free to fluorescence?

[mattersoflight]

@yingmuzhi glad to hear that our data and code have been helpful. I suspect the problem you are seeing is due to normalization per FOV. When virtual staining across a large structure, it is important to normalize across the whole dataset. See this result: https://elifesciences.org/articles/55502/figures#fig7s1. There is a corresponding paragraph in the methods section of the paper that describes the details of the normalization - the main detail is that one should use median and inter-quartile range, rather than mean and standard deviation for z-scoring the dataset.

If you post your config and some example images, we can comment further.

[yingmuzhi]

Hello, professor Mehta. I am Very Glad that you reached out. I have seen the result https://elifesciences.org/articles/55502/figures#fig7s1 and I used median and IQR across the dataset(although is a mini dataset from the whole dataset), not per FOV. Median and IQR works well to address the issue of background distortion when I was doing the cell experiment and I kept on using it. But I can not see axon tracks clearly.

down below are some experiments details, hoping you can help me figure this problem,

• preprocess.yml https://github.com/yingmuzhi/issues/blob/master/issue01/preprocess.yml
• train.yml https://github.com/yingmuzhi/issues/blob/master/issue01/train.yml
• inference.yml https://github.com/yingmuzhi/issues/blob/master/issue01/inference.yml
• model_graph.png https://github.com/yingmuzhi/issues/blob/master/issue01/model_graph.png
• history.csv https://github.com/yingmuzhi/issues/blob/master/issue01/history.csv
• metrics_xy.csv https://github.com/yingmuzhi/issues/blob/master/issue01/metrics_xy.csv
• result_one_pic.png https://github.com/yingmuzhi/issues/blob/master/issue01/img_ex561em700_t000_p016_z000.jpg
• input.tif https://github.com/yingmuzhi/issues/blob/master/issue01/signal_t000_p016_z000.png
• prediction.tif https://github.com/yingmuzhi/issues/blob/master/issue01/prediction_t000_p016_z000.png I changed the grayscale histogram since the origin image is too dark.
• label.tif https://github.com/yingmuzhi/issues/blob/master/issue01/target_t000_p016_z000.png

[ziw-liu]

@yingmuzhi I can't access the files you linked. Can you double check that they are valid?

[yingmuzhi]

@ziw-liu hello, doctor Liu, I tried copy the url like https://github.com/yingmuzhi/issues/blob/master/issue01/preprocess.yml in the search bar, and then it will show up, not click directly.