Latent Space Oversampling for Addressing Adult-Pediatric Imbalance in Brain Tumor Segmentation

Saeideh Ghanbari Azar, Tufve Nyholm, Tommy Löfstedt

Keywords: brain tumor segmentation, latent-space oversampling, adult–pediatric imbalance, StyleGAN2, SMOTE

TL;DR: A method for reducing adult–pediatric performance gaps in brain tumor segmentation using latent-space oversampling with a cohort-conditioned StyleGAN2.

Abstract: Deep-learning models tend to perform unevenly when there are imbalances in training sample subgroups, e.g., in the number of adult and pediatric training samples in brain tumor segmentation. There are often fewer pediatric scans and they differ from adult scans in anatomy, contrast, and tumor characteristics. In this work, we studied if enriching the pediatric cohort with realistic synthetic samples would resolve this imbalance. Specifically, we sought to know if latent space minority oversampling methods would resolve the imbalance and whether the quality of the latent space would make a difference in this application. We first constructed an adult-pediatric dataset by unifying the BraTS and BraTS-PEDs datasets, making them compatible through consistent preprocessing and labeling. We then developed a cohort-conditioned StyleGAN2 model to jointly model multi-modal MRI slices and tumor masks. Pediatric slices were embedded into the generator's latent space and, using the Synthetic Minority Over-sampling Technique (SMOTE), new pediatric latent vectors were produced. These new latent vectors were decoded into MRI-mask sets and added to the training set to balance the adult-pediatric counts. This proposed latent space oversampling strategy was compared to several imbalance-mitigation baselines. Evaluations on a balanced test set of 200 adult and 200 pediatric subjects showed that the proposed latent space oversampling improves pediatric Dice scores without decreasing the adult performance and obtains the smallest adult-pediatric performance gap of all evaluated methods.

Primary Subject Area: Fairness and Bias

Secondary Subject Area: Image Synthesis

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 228