Image size when training with Glas dataset about medical-transformer

[jeya-maria-jose]

During evaluation, we resize the ground truth also to 128*128 resolution to match the predictions.

[jeya-maria-jose]

Hi,

Thanks for your comment. We resized all the images to 128x128 for our experiment as well as baseline comparisons. I agree that training on a higher resolution might be more helpful and result in a better performance. As we just focused on making the case for MedT, we just followed a standard resolution for all the experiments. We can definitely analyse doing patch wise training for MedT as well as other baselines for GlaS in the future to further push the performance.

[weiaicunzai]

Hi,

Thanks for your comment. We resized all the images to 128x128 for our experiment as well as baseline comparisons. I agree that training on a higher resolution might be more helpful and result in a better performance. As we just focused on making the case for MedT, we just followed a standard resolution for all the experiments. We can definitely analyse doing patch wise training for MedT as well as other baselines for GlaS in the future to further push the performance.

Thanks for your replay, just want to make sure, during the evaluation phase, you resize the prediction of network from 128 * 128 to the original image size?