Handling Images without Tumors for Model Training and Testing

med-air / 3DSAM-adapter

Holistic Adaptation of SAM from 2D to 3D for Promptable Medical Image Segmentation

131 stars 12 forks source link

Handling Images without Tumors for Model Training and Testing #35

Open zhou07 opened 3 months ago

zhou07 commented 3 months ago

Hi,

Great work! I have a question. Some images in my dataset do not contain tumors, so their masks are completely zero. How can prompt points be generated for these images during training and testing?

peterant330 commented 2 months ago

Hi,

During the training we randomly pick points from the background for these images.

For testing, we did not include such cases for calculating the metrics. But for real application, it would have some tolerance if you click a background point and no masks will be generated.