Abstract
Deep learning models tend to perform unevenly when there are imbalances in label frequencies or sample subgroups. In this study, we investigated adult–pediatric imbalance in brain tumor segmentation. Pediatric data usually differ from adult data in their number of samples, anatomy, contrast, noise, and tumor characteristics. Since there are no datasets containing both adult and pediatric medical images specifically designed to investigate this problem, we constructed an imbalanced adult–pediatric dataset by unifying the BraTS and BraTS-PEDs datasets, using a consistent pipeline to preprocess and harmonize them, resulting in a reusable benchmark for studies on imbalanced data. Solutions to imbalance problems often rely on enriching the minority group with realistic synthetic samples. However, existing methods are typically not designed for unstructured data such as images. We therefore investigated whether generative models can be used to address adult– pediatric imbalances and, more specifically, whether latent space oversampling methods such as DeepSMOTE can benefit from more expressive latent representations. To this end, we propose a methodological extension of DeepSMOTE that replaces its autoencoder latent space with the latent space of a cohort-conditioned StyleGAN2 model and evaluated it on naturally imbalanced adult and pediatric brain tumor images. The pediatric slices were embedded into the StyleGAN2 latent space and, using Synthetic Minority Over-sampling Technique (SMOTE), new latent vectors were produced and decoded into MRI–mask pairs to balance the number of adult and pediatric samples. The proposed latent space oversampling strategy was compared to several baseline methods. Evaluations on a balanced test set of 200 adult and 200 pediatric subjects showed that the proposed StyleGAN2 latent-space method improves pediatric Dice scores without decreasing adult performance, outperforms the autoencoder latent-space variant, and achieves the smallest adult– pediatric performance gap among all evaluated methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/PedAItrics_006.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{GhaSae_Latent_MICCAISAT2026,
author = { Ghanbari Azar, Saeideh AND Nyholm, Tufve AND Löfstedt, Tommy},
title = { { Latent Space Oversampling for Adult–Pediatric Imbalance in Brain Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17257},
month = {pending},
page = {pending}
}