any suggestions on how to adapt the model to work on dataset with much smaller tumors or areas of interest?

[wshi8]

Your advice is appreciated.

[askerlee]

You can take a two-stage approach: first train a model to do localization only. Then feed the cropped images into a segmentation model like segtran or unet, and the images needn't to be downsampled as they are already small.

[wshi8]

Thank you!