search

MIC-DKFZ / MultiTalent

Implemention of the Paper "MultiTalent: A Multi-Dataset Approach to Medical Image Segmentation"
Apache License 2.0
49 stars 3 forks source link

Question about Pretrained MultiTalent Model's Handling of Lung/Lung Tumor Labels #11

Closed Awayah closed 9 months ago

Awayah commented 9 months ago

HI, I recently ran inference on a lung tumor dataset using the pretrained MultiTalent model. Overall, the performance on the labeled tumor data 06_lungnodule was promising.

However, when I was examining left and right lung tissue segmentation outputs, (51_left_lung, 51_right_lung) I observed the presence of 'holes' in the segmentation corresponding to the tumor regions, particularly noticeable with larger tumors. My question is whether the original dataset used for training (specifically Dataset ID: 51) included labeled lung tissue with or without the pathology.

I haven't been able to get access to the relevant dataset used for pretraining (ID: 51), so I would greatly appreciate clarification on it.

Thank you, Awayah

Example:

image
constantinulrich commented 9 months ago

Hi, dataset 51 originates from the structsec challenge (task3) that is about the segmentation of organs at risk. Therefore, I assume that they did not include the target tumors in the segmentation masks of the OAR...

constantinulrich commented 9 months ago

I checked one of the training images: They exclude the tumor... image

Awayah commented 9 months ago

Thank you for verifying!