By directing each MRI scan to a specialised model depending on whether the patient has undergone surgery or not, the team has improved the Dice coefficient by more than 4 points on a lesion-by-lesion basis for post-treatment cases.
The BraTS-GLI 2025 challenge required segmenting glimes on multimodal MRIs, both before and after treatment. The difficulty: the cavity left by surgical excision exists only in patients already operated, so that the two groups do not have the same targets to define. One team designed a variant of nnU-Net v2 that explicitly manages this asymmetry.
A referral rather than a set
During joint training, the loss function causes a controlled penalty for false detections on empty channels, typically the operating cavity in an unoperated patient. At inference, each examination is directed to a single specialized model according to treatment status, without the use of synthetic tumours, votes or sets of models.
Ranked second, the method obtains on the blind test set Dice scores by lesion of 0.79 for the enhanced tumor, 0.87 for the cavity, 0.79 for the tumour heart and 0.87 for the entire tumor. On post-operative cases, taking into account the cavity improves the average Dice by 0.042 compared to a conventional common drive.
Read the full article on the original source
Go to the source →