Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-calibrated convolution towards glioma segmentation

Felipe C. R. Salvagnini, Gerson O. Barbosa, Alexandre X. Falcao, Cid A. N. Santos

TL;DR

The study addresses automated glioma segmentation across multi-modal MRI by integrating self-calibrated convolutions (SC-Conv) into the nnU-Net framework. Through experiments on BraTS 2023 data, the authors test SC-Conv placements in encoder, decoder, and skip connections, finding that SC-Conv in skip connections yields the best improvement in enhanced-tumor and tumor-core segmentation while keeping whole-tumor performance stable. Although costs vary by placement, the skip-connection configuration demonstrates the potential of lightweight, locality-aware context modeling to enhance clinically relevant regions. This work highlights SC-Conv as a practical augmentation to established multi-modal segmentation pipelines, with implications for robust cross-institution brain tumor delineation.

Abstract

Accurate brain tumor segmentation in the early stages of the disease is crucial for the treatment's effectiveness, avoiding exhaustive visual inspection of a qualified specialist on 3D MR brain images of multiple protocols (e.g., T1, T2, T2-FLAIR, T1-Gd). Several networks exist for Glioma segmentation, being nnU-Net one of the best. In this work, we evaluate self-calibrated convolutions in different parts of the nnU-Net network to demonstrate that self-calibrated modules in skip connections can significantly improve the enhanced-tumor and tumor-core segmentation accuracy while preserving the wholetumor segmentation accuracy.

Self-calibrated convolution towards glioma segmentation

TL;DR

The study addresses automated glioma segmentation across multi-modal MRI by integrating self-calibrated convolutions (SC-Conv) into the nnU-Net framework. Through experiments on BraTS 2023 data, the authors test SC-Conv placements in encoder, decoder, and skip connections, finding that SC-Conv in skip connections yields the best improvement in enhanced-tumor and tumor-core segmentation while keeping whole-tumor performance stable. Although costs vary by placement, the skip-connection configuration demonstrates the potential of lightweight, locality-aware context modeling to enhance clinically relevant regions. This work highlights SC-Conv as a practical augmentation to established multi-modal segmentation pipelines, with implications for robust cross-institution brain tumor delineation.

Abstract

Accurate brain tumor segmentation in the early stages of the disease is crucial for the treatment's effectiveness, avoiding exhaustive visual inspection of a qualified specialist on 3D MR brain images of multiple protocols (e.g., T1, T2, T2-FLAIR, T1-Gd). Several networks exist for Glioma segmentation, being nnU-Net one of the best. In this work, we evaluate self-calibrated convolutions in different parts of the nnU-Net network to demonstrate that self-calibrated modules in skip connections can significantly improve the enhanced-tumor and tumor-core segmentation accuracy while preserving the wholetumor segmentation accuracy.
Paper Structure (10 sections, 4 figures, 1 table)

This paper contains 10 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adding SC-Conv into U-shaped Networks
  4. Experimental setup
  5. Dataset
  6. Network Structure
  7. Training procedure
  8. Evaluation metric
  9. Results
  10. Conclusion

Figures (4)

  • Figure 1: 3D SC-Conv Module self_calibrated.
  • Figure 2: Data instance from BraTS23. One can visualize NCR (hypointense) and ET (hyperintense) tissues on T1Gd, and ED (hyperintense) on T2-FLAIR protocols.
  • Figure 3: Best architecture found (nnU-Netnnunet with SC-Conv on skip-connections).
  • Figure 4: Results for BraTS subject 00019. The first column shows T1-Gd images, and the second one, T2-Flair. a-b are results with SC-Conv on skip-connections, c-d the ground truth, and e-f the baseline's result.