CTransPath WSInfer
Hi Jakub,
Greetings from Zheng. WSInfer is a extremely useful tool for my study. I saw you share the torchscript of CTransPath in your model. May I know if the CTransPath model could be loaded on WSInfer QuPath extension? At the same time, our studies are focusing on human peripheral blood cells. I'd like to know if there have any model that could be utilized in WSInfer QuPath extension to recognize such as neutrophils, eosinophils, plasma cells,etc? Many thanks for you and I looking forward to hearing any advice from you.
hi @masudalynn - ctranspath is a patch encoder. it embeds an H&E image patch to a vector with 768 dimensions. it doesn't classify patches, so we don't include it in wsinfer.
for segmenting cell types, you can look into HoverNet: https://github.com/vqdang/hover_net . it can't be used in wsinfer, but their code has inference scripts for tiles and whole slide images. though hovernet doesn't distinguish between different immune cells. it lumps all immune cells into one label.
let me know if you find a model that classifies between different immune cells. i would be interested to try it out
Hi Jakub,
Many thanks for your prompt reply. I deeply interested in some types of immune cells especially neutrophils and eosinophils. I found a deep learning model named EfficientNet-Peripheral-Blood-Cell generated by Sarah Antillia (https://github.com/sarah-antillia/EfficientNet-Peripheral-Blood-Cell/tree/main?tab=readme-ov-file). May I know if this model could be used in wsinfer? Thanks again for your help.
yes that model could be used in wsinfer. you can create a torchscript version of the trained model, and then use that with wsinfer, either through the command line or qupath.
however, that model expects the cell to be in the center of the image. wsinfer can't guarantee that the cells of interest will be in the center of each patch. you could record the points of cells and then extract patches, so that the cell of interest is in the center.