Task1Baseline
About
Summary
Task 1 Baseline
Mechanism
My original idea for Task 1: Segmentation on diagnostic MRIs. Utilize nnU-Net’s Residual Encoder UNets (ResEncM) to perform 5-fold supervised learning on 92 labeled samples, 5-fold soft ensemble pseudo-labeling on 367 unlabeled samples, and 5-fold semi-supervised learning on 389 background-only-excluded samples.
However, this strategy does not provide a competitive performance to significantly outperform the baseline method. Therefore, I return back to the baseline approach which utilizes 3-fold nnU-Net modules to perform the binary segmentation.
For post-processing, I tried to leverage all 5 folds of the MRSegmentator, but that did not lead to a better performance in the open development phase of Task 1, while increasing the running time for each test case. Therefore, when submitting for the closed testting phase of Task 1, I only activated 1 MRSegmentator checkpoint.
Currently, the activated container image and model are implementations of my original idea for Task 1, instead of the version I submitted for Task 1's cloased testing phase, as that would be exactly the same as the baseline method.
Thank you for your time and reading.
Interfaces
This algorithm implements all of the following input-output combinations:
Validation and Performance
Challenge Performance
| Date | Challenge | Phase | Rank |
|---|---|---|---|
| June 22, 2025 | PANTHER | Open Development Phase Task 1 | 54 |
| Sept. 7, 2025 | PANTHER | Closed Testing Phase Task 1 | 10 |
Uses and Directions
This algorithm was developed for research purposes only.