Abstract

Accurate segmentation of pediatric brain tumor subregions in multi-parametric magnetic resonance imaging is challenging because tumor appearance varies across cases and small enhancing tumor (ET) and cystic component (CC) regions may contain few voxels. We present a task-specific adaptation of PTransBTS for BraTS-PEDs 2026 Task 2. The submitted configuration combines modality-specific stems, a convolution-only backbone path, inherited attention-based pyramid priors, five-class prediction, Dice-Focal deep supervision, and extensive spatial and intensity augmentation. At inference, logits are averaged across eight flip views within each fold model, followed by softmax and probability averaging across five models; threshold-based class-volume filtering and geometry restoration produce the final NIfTI segmentation. On the official Synapse validation set (submission 9772538), global DSCs were 0.5096, 0.9085, 0.2196, 0.0090, 0.9400, and 0.9415 for ET, NET, CC, ED, TC, and WT, respectively. These results show substantially higher overlap for composite regions than for fine-grained classes, with ED and CC remaining the main limitations. Source code is available at https://github.com/hxc236/PTransBTS-PED.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_PEDs_010.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=uPPpU4ZkWS

BibTex

@InProceedings{HouXia_Adapting_MICCAISAT2026,
        author = { Hou, Xiangchen AND Peng, Yanjun AND Yu, Haitao},
        title = { { Adapting PTransBTS for Pediatric Brain Tumor Segmentation in BraTS-PEDs 2026 } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17253},
        month = {pending},
        page = {pending}
}


back to top