Benchmark & Code

In the iSeg-2019 review article [1], “Multi-Site Infant Brain Segmentation Algorithms: The iSeg-2019 Challenge”, IEEE Transactions on Medical Imaging, 2021, we draw the following major conclusions:

1. The multi-site issue that learning-based models often perform poorly on testing subjects acquired with different imaging protocols/scanners as the training subjects, hinders the popularity and practicability of learning-based methods.
2. Although most participating methods employed advanced deep learning techniques, none of them can achieve consistent performance across different sites with different scanners/imaging protocols.
3. The multi-site issue is mainly caused by the different scanner/imaging protocols. It might be worth investigating to harmonize images from different sites based on MR imaging physics.
4. Domain adaptation is a possible way to alleviate the multi-site issue.
5. It might be worth exploring the site-independent anatomy prior information [2] to alleviate the multi-site issue.

The multi-site issue is still an open question and the iSeg-2019 website is always open.

[1]Yue Sun, et al., “Multi-Site Infant Brain Segmentation Algorithms: The iSeg-2019 Challenge,” in IEEE Transactions on Medical Imaging, 40(5), 1363-1376, 2021.

[2]Li Wang, et al., “Volume-Based Analysis of 6-Month-Old Infant Brain MRI for Autism Biomarker Identification and Early Diagnosis,” in MICCAI, 2018, pp. 411-419. [PDF] [Code: Caffe prototxt] [Software]

If you use the iSeg-2019 datasets and like to post your publication/code in this page, please kindly contact li_wang@med.unc.edu.