Abstract

Pediatric brain tumor segmentation from multiparametric MRI remains challenging due to severe class imbalance and the rarity of certain subregions such as cystic component and peritumoral edema. We present an ensembling framework built upon nnU-Net and MedNeXt for the BraTS-PEDs 2026 Challenge, targeting the four-label system comprising enhancing tumor (ET), non-enhancing tumor core (NET/C), cystic component (CC), and peritumoral edema (ED). Track 1 handles ET, NET/C, and CC through class-specific pAbstract. We present YOLO-UNet, a 3D segmentation network for the BraTS-PEDs 2026 challenge (Task 2) that couples a YOLO-style encoder—a CSPDarknet backbone with a path-aggregation feature pyra mid (PAFPN)—to a U-Net-style decoder with skip connections. On top of this backbone we introduce a full-resolution residual boundary branch: a lightweight module that extracts high-frequency features directly from the native-resolution input image and adds a zero-initialized residual correction to the decoder logits, which are otherwise produced at re duced resolution and trilinearly upsampled. Because the correction head is zero-initialized, the model output is identical to the interpolation limited baseline at initialization and subsequently learns only a residual refinement. We further use a region-selective ensemble of ten models: five full-resolution residual models supply the whole-tumor and tumor core channels, while five base models supply the enhancing-tumor and cystic channels. On the official validation set (91 cases) our method at tains Dice scores of 0.940 (whole tumor), 0.939 (tumor core), 0.910 (non enhancing tumor core) and 0.577 (enhancing tumor), with a normalized surface Dice of 0.664 for whole tumor. Each individual model is com pact at 18.95M parameters, although the submitted system ensembles ten of them under four-view test-time augmentation. Code is available at https://github.com/tshzg/-yolo-unet-brats-peds.robability fusion over an eightmodel nnU-Net ensemble, where snapshot selection monitors per-class Dice trajectories to capture subregion-specialist checkpoints. Track 2 handles the substantially harder ED label through a strict three-way voxel-wise intersection over hard segmentation masks from two Med- NeXt models and one nnU-Net checkpoint selected for the ED consensus branch, conservatively predicting ED only where all three models agree and applying the three-model ED consensus mask as a post-argmax override. On the BraTS-PEDs 2026 validation set of 91 cases, the method achieves lesion-wise Dice Similarity Coefficients of 0.587, 0.922, 0.277, and 0.026 for ET, NET/C, CC, and ED respectively, with Tumor Core and Whole Tumor DSC of 0.950 each. The dedicated ED branch achieved an ED DSC of 0.026, compared with 0.012 for the MedNeXt soft-voting baseline.



Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_PEDs_003.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=E1fmhAV9jb

BibTex

@InProceedings{TanSha_YOLOUNet_MICCAISAT2026,
        author = { Tang, Shaohua AND Jung, Young-Jin},
        title = { { YOLO-UNet: A Hybrid YOLO-Encoder / U-Net-Decoder Network with a Full-Resolution Residual Boundary Branch for Pediatric Brain Tumor Segmentation } },
        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