Can't get satisfactory results on kidney tumor segmentation

[mlszy928]

Hi, thank you for the great work! I trained the MedSAM-adapter on 2D 256x256 CT images obtained from Kidney Tumor Segmentation (KiTS) dataset for 50 epoches, the best model is achieved on epoch 11. But the DSC score of the testing data is not more than 0.7, but it is 0.76 from MedSAM. How should I do for the dataset? If I should reduce the size of the model?

By the way, I have changed the proj layer to fit the size of my dataset as following:

net.image_encoder.patch_embed.proj = nn.Conv2d( 3, 768, kernel_size=4, stride=4, padding=0 )