Work

MSPE-Swin: Multi-Scale Patch Embedding for Robust SwinUNETR — Weights

Year: 2026

Type: dataset

Abstract: Pretrained model weights for MSPE-Swin: Multi-Scale Patch Embedding For Robust SwinUNETR. MSPE-Swin adapts Multi-Scale Patch Embedding to SwinUNETR by replacing only the patch embedding layer with a small set of resolution-specific patch kernels, leaving the encoder, the skip connections, and the decoder unchanged. This record contains the checkpoints for the pulmonary-vessel histology, ADE20K, and HVSMR-2.0 experiments, under both the frozen-backbone fine-tuning and the full mixed-resolution training regimes, together with the ablation checkpoints. Source code: https://github.com/Digiratory/SwinUNETR-MSPE

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.21526767

Authors: Lonchakov, Aleksandr, Kaplun, Dmitry, Sinitca, Aleksandr

Institution: Saint Petersburg State Electrotechnical University

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7170252082

Cites:

Cited by:

Related to: 10