search

uni-medical / SAM-Med3D

SAM-Med3D: An Efficient General-purpose Promptable Segmentation Model for 3D Volumetric Medical Image
Apache License 2.0
531 stars 71 forks source link

【train】【performance】about the fine tune result #61

Closed SillyPigeon closed 2 months ago

SillyPigeon commented 7 months ago

Hi Thanks for your great jobs. However, when I try to fine tune dataset in my vessels data(for train 200+epochs in imageCAS dataset). The inference result was so bad that it could not be used in project. Do you have some ideas about this?

image

the main picture is GT

image
blueyo0 commented 2 months ago

Hi, Coronary Artery is a challenging target, and I suggest you adjust the window size (instead of 128, you can use a larger window to make sure the artery is continuous) and spacing of the data (instead of 1.5, maybe using sparse slices with high-resolution images 3x0.8x0.8 is better?). Anyway, specific modification of data processing seems help for your task. Thx.