[Paper Review] Brain Tumor Segmentation with Special Emphasis on the Non-Enhancing Brain Tumor Compartment

T. Schaffer, A. Brawanski+3 more|arXiv (Cornell University)|Feb 25, 2026

Brain Tumor Detection and Classification0 citations

TL;DR

The paper introduces an Upscaling PAU-Net for brain tumor segmentation that explicitly includes the non-enhancing tumor (NET) compartment, achieving competitive Dice scores on BraTS 2018/2021 and enabling four-label (ET, ED, NCR, NET) and later four-label with NET extension.

ABSTRACT

A U-Net based deep learning architecture is designed to segment brain tumors as they appear on various MRI modalities. Special emphasis is lent to the non-enhancing tumor compartment. The latter has not been considered anymore in recent brain tumor segmentation challenges like the MICCAI challenges. However, it is considered to be indicative of the survival time of the patient as well as of areas of further tumor growth. Hence it deems essential to have means to automatically delineate its extension within the tumor.

Motivation & Objective

• Motivate accurate segmentation of all four tumor compartments, with special emphasis on the non-enhancing tumor (NET) region.
• Develop a high-resolution, upscaling U-Net variant (PAU-Net) to improve NET delineation.
• Evaluate multiple residual filter block architectures and an upscaling decoder to balance performance and overfitting.
• Extract NET compartments from BraTS datasets by leveraging a trained NET predictor and morphological filtering.
• Create unified 4-label BraTS datasets (NCR, ED, NET, ET) and assess segmentation performance across BraTS challenges.

Proposed method

• Adopt a 3D U-Net-like encoder-decoder architecture with four resolution levels.
• Experiment with four base residual filter block variants (original, ResNet-like, pre-activation, and pre-activation with post-activation differences).
• Develop an upscaling PAU-Net by adding a high-resolution decoder branch to produce segmentation masks at twice the resolution.
• Upscale ground-truth NET masks and train with soft Dice loss; evaluate with Dice-Sørensen, IoU, and Hausdorff Distance.
• For NET extraction on BraTS 2018, decompose NCR/NET using a BraTS 2021-trained PAU-Net to predict NET and apply morphological filtering to clean masks.
• Extend the unified BraTS 2018/2021 dataset to a four-label scheme (ET, TC, WT, NET) and train a 4-label PAU-Net on cropped inputs (4 × 96 × 192 × 160).

Figure 1: Different types of residual filter blocks. Each type can be used in the above U-Net architecture

Experimental results

Research questions

• RQ1 Can a high-resolution PAU-Net effectively segment the non-enhancing tumor (NET) compartment in brain MRI?
• RQ2 Do residual filter block variants and an upscaling decoder improve segmentation performance on the ET, TC, WT, and NET compartments compared to standard U-Net baselines?
• RQ3 Is it feasible to derive NET masks from BraTS datasets lacking explicit NET labels using a NET-predictive model and morphological filtering?
• RQ4 What are the comparative Dice scores across BraTS 2018 and BraTS 2021 when training on a unified four-label BraTS dataset?
• RQ5 Does a 4-level (vs 5-level) PAU-Net generalize better to test data while maintaining competitive training performance?

Key findings

• The 4-level upscaling PAU-Net generally outperformed the 5-level variant on BraTS 2021 test data (and showed competitive results on BraTS 2018 test data).
• On BraTS 2018 train/test, ET: 0.7179/0.7802, TC: 0.8646/0.8230, WT: 0.9010/0.8912 (4-level model often comparable or superior to 5-level).
• On BraTS 2021 train/test, ET: 0.8566/0.8464, TC: 0.9121/0.8841, WT: 0.9232/0.9186 (4-level model often slightly better in test).
• Mean Dice across datasets favors the 4-level PAU-Net in test scenarios (e.g., BraTS 2021 test mean: 0.8830 for 4-level vs 0.8771 for 5-level).
• Extraction of NET from BraTS 2018 was achieved by decomposing NCR masks using a BraTS 2021-trained NET predictor and applying morphological filtering to clean results.
• A unified 4-label BraTS 2018/2021 dataset (NCR, ED, NET, ET) was created and used to train a four-label PAU-Net with a cropped input size, achieving Dice scores across compartments.

Figure 2: Two variants of the new up-scaling PAU-Net.