Master Assignment
BABY BRAINS 3: Development Aware Pediatric Brain MRI Segmentation
Type: Master CS
Student: Unassigned
Duration: TBD
If you are interested please contact:
Background:
Foundation Models are everywhere around us, and taking up an increasing role in medicine, helping towards earlier and more accurate diagnoses, personalized care and interventions, and reducing clinicians’ burnout.
However, although 26.8 % of humans are children, only 2.2% of AI research is pediatric. And children are not mini-adults: their data will not follow the same patterns as that of adults. Moreover, kids are not mini-adults and kids vary a lot at different developmental stages. As a result, the AI solutions developed and trained on adult data underperform on children, as concluded in a 2026 scoping review [1]. A very new paper from WACV 2027 [2] tested foundation models on chest X-rays from children and found performance drops of macro-AUC from 0.92 to 0.78.
The drop in performance can be attributed to a variety of factors, related both to the data and the models and methods used. There are a few benchmarking datasets of brain MRIS from babies and children, described below, that should be used in the topics below. They have been chosen purposefully, for consistency of data type and annotations, but still feature distribution shifts related to age, hospital and other factors.
This project aims to adapt medical image foundation models according to infants’ or childrens’ age, to achieve robust brain area segmentation. It asks the question if the inclusion of developmental age can improve segmentation results, first by comparing them by age, and then by finding how to improve performance for different age groups.
Objectives:
· Quantify age-related variation in segmentation performance.
· Develop a lightweight age-conditioned adapter (e.g. using PEFT fine tuning).
· Compare age-aware and age-independent adapters at matched capacity.
· Evaluate age-specific results and perform age-shuffling ablations.
Baby Brains Benchmark Datasets
The datasets below should be used, in the order presented. The reason is these are datasets with consistent labeling and modality. You should examine both brain segmentation as a first easy step. You should then examine the segmentation of finer details, which will enable the comparison of different models and tuning approaches. The level of detail should be limited to White Matter (WM), Gray Matter (GM) and CerebroSpinal Fluid (CSF).
References:
- Small sample, uneven age variation, but most practical for first experiments.
- Includes curated segmentations across different infant ages, making it attractive for studying developmental changes using a consistent dataset. It is smaller than HBCD, but its annotations make it a more manageable initial experiment.
2) iSeg 2017
- T1w and T2w MRI with manually annotated WM, GM, and CSF. The benchmark used 10 training subjects and 13 hidden-test subjects, all approximately six months old. Excellent for testing whether MedSAM-like models or pretrained encoders can be adapted to infant brain segmentation.
- One age stage, cannot be used to study over
3) iSeg 2019
- new version of iSeg 2017, also one age (about 10 months)
- one training data acquisition site, three different test sites (vs one in 2017)
4) HEALthy Brain and Child Development Study: you have to request access for this data.
MRI visits at 0–1, 3–9, and 9–15 months. Strong for evaluating age-dependent performance.
5) Laborie LB, Lee J, Antram E, Lein RK, Shelmerdine SC. Performance of adult-trained artificial intelligence models in paediatric imaging-a scoping review. Eur Radiol. 2026;36(7):5916-5939. doi:10.1007/s00330-026-12354-5.
6) Sara Inácio, Hugo Proença, João C. Neves, From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift, accepted WACV 2027, October 2026, arXiv:2610.21809.
7) Chen JV, Li Y, Tang F, et al. Automated neonatal nnU-Net brain MRI extractor trained on a large multi-institutional dataset. Sci Rep. 2024;14(1):4583. Published 2024 Feb 26. doi:10.1038/s41598-024-54436-8
8) Timothy J. Hendrickson, et al, BIBSNet: A deep learning baby image brain segmentation network for MRI scans, Developmental Cognitive Neuroscience, Volume 79, 2026, 101706, ISSN 1878-9293, https://doi.org/10.1016/j.dcn.2026.101706.
9) Sun, Z., Huang, J., Ma, X. et al. A Low-Field MRI Dataset For Spatiotemporal Analysis of Developing Brain. Sci Data 12, 109 (2025). https://doi.org/10.1038/s41597-025-04450-w.
10) A. K. Shibu, S. Ramanarayanan, V. Kanna, J. Jayakumar, K. Ram and M. Sivaprakasam, MedSAM-Guided Curriculum Learning for White Matter Tract Segmentation in Block Face Imaging of Fetal Brain, 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), Honolulu, HI, USA, 2025, pp. 1045-1052, doi: 10.1109/ICCVW69036.2025.00113.
11) Zeng, Q., Liu, W., Li, B., Didier, R., Grant, P.E., & Karimi, D. (2026). Atlas-Assisted Segment Anything Model for Fetal Brain MRI (FeTal-SAM). ArXiv, abs/2601.15759.