Questions about TCGA SVS Reproduction & Cell Prediction Optimization + Proposal for ipynb Script

[easyeryiji-cyber]

Dear Authors,

First of all, I would like to express my sincere admiration and gratitude for your groundbreaking work on GigaTIME! It's truly impressive how you bridge H&E and mIF images, and your open-sourced code has been extremely helpful for my research.

I have successfully deployed your code and reproduced the workflow on TCGA SVS whole-slide images (WSIs) (directly downloaded from the TCGA database). The overall pipeline is runnable without critical errors, which is great news. However, I encountered some minor issues with cell prediction quality and have a few questions to seek your valuable advice:

1. About the prediction cutoff value: Is the recommended binarization cutoff still around 0.5 (i.e., pred = (probs > 0.5).float()) for cell type prediction? I used this value but wonder if it needs adjustment for TCGA data (considering potential differences in scanning quality or tissue types).

2. About TCGA SVS layer selection: The SVS files downloaded from TCGA have 4 resolution levels (indexed from 0 to 3). My WSI resolution details are as follows:

Seeded. Torch: 2.10.0, CUDA: Not available, MPS Available: True, Device: mps
MPS Acceleration Enabled (PyTorch 2.10.0 built-in MPS operators are supported to optimize inference performance)
=> Successfully loaded model: gigatime (MPS inference mode is enabled and optimized)
=> Cured channel list: 23 channels in total, with 21 valid channels (background channels excluded)

SVS Whole Slide Image (WSI) Information:
Dimensions (Width × Height): 103910 × 79454
Number of resolution levels: 4
Dimensions of each level: ((103910, 79454), (25977, 19863), (6494, 4965), (3247, 2482))
Total number of generated patches: 31668

I used the 1st level (index 0, the lowest resolution with the highest downsampling) for my experiments. The 1st level (index 0) is the highest resolution (103910 × 79454) without downsampling. Is my layer selection (index 0) the correct choice for GigaTIME?

3. About cell alignment and prediction quality: I tiled the SVS into 512×512 patches and used an overlapping sliding window strategy to mitigate cell misalignment during prediction. However, the final results are not as perfect as your demo data (cell boundaries are not as clear, and there's slight misalignment in some regions). Do you have any specific suggestions to optimize the results for TCGA SVS data?

Additionally, I noticed that there is currently no dedicated tutorial or script for reproducing GigaTIME on TCGA SVS data. For users like me who want to apply your work to TCGA datasets efficiently, a well-documented Jupyter Notebook (ipynb) script would be extremely helpful.

Would it be possible for you to upload an official ipynb script (which could include steps like SVS layer loading, 512×512 patching, sliding window inference, and result visualization) to the project's scripts/ directory (or a new tcga_demo/ directory)? This would greatly lower the barrier for other researchers to reproduce and extend GigaTIME on TCGA data.

Thank you very much for your time and guidance. Looking forward to your reply!

Best regards,

Ji

[jingyu9603]

@easyeryiji-cyber Dear，

I came across your thoughtful request regarding the need for a dedicated tutorial or script to run GigaTIME on TCGA SVS data. I am currently facing the exact same challenge and completely agree that a well-documented Jupyter Notebook would be immensely helpful for researchers like us.

I was wondering if you have made any progress in addressing this issue since your post. If you have managed to develop or obtain a working script (e.g., for SVS loading, patching, inference, and visualization), would you be willing to share it? Any guidance or code snippets you could provide would be greatly appreciated and would certainly help lower the barrier for others in the community.

Thank you very much.