Efficiently correcting patch-based segmentation errors to control image-level performance

Experiments for MIDL 2024 submission.

Instructions to Reproduce the Results:

1. Download FIVES dataset (Jin et al., 2022) and register its path in paths.yaml.
2. Execute 0_pass_forward.ipynb to pass the data through the ensemble and log the outputs. The models are already trained and their weights are stored in trained/.
3. Calibrate with temperature scaling on the validation set with 1_calibrate.ipynb
4. Compute $\widehat{DSC}$ and $\widehat{DSC}_{\text{corr}}$ in 2_estimate_DSC.ipynb
5. Reproduce figures with respective scripts

References: Jin, Kai, et al. "Fives: A fundus image dataset for artificial Intelligence based vessel segmentation." Scientific Data 9.1 (2022): 475.