Hi, I've tested DeepSpot-M in a dataset with co-registered COSMX WTx and H&E pancreas data. I've obtained quite bad/low pearson correlation, RMSE and correlation of Moran's I values between predicted vs original expression values, compared to similar other tools I have used for gene expression prediction from H&E. I have used DeepSpot in the same dataset, and performance with the trained model on my own data, was way better. I followed the same settings in the example codes in this GitHub.
Are there any recommendations to improve performance?
How can I perform test-time adaptation on my samples (e.g., in one- or few-shot mode), or fine-tune the model with my own data? This does not seem to be disclosed in the GitHub tutorials, even though mentioned in the paper.
Thanks!
Hi, I've tested DeepSpot-M in a dataset with co-registered COSMX WTx and H&E pancreas data. I've obtained quite bad/low pearson correlation, RMSE and correlation of Moran's I values between predicted vs original expression values, compared to similar other tools I have used for gene expression prediction from H&E. I have used DeepSpot in the same dataset, and performance with the trained model on my own data, was way better. I followed the same settings in the example codes in this GitHub.
Are there any recommendations to improve performance?
How can I perform test-time adaptation on my samples (e.g., in one- or few-shot mode), or fine-tune the model with my own data? This does not seem to be disclosed in the GitHub tutorials, even though mentioned in the paper.
Thanks!