Brain Tumor Segmentation with Special Emphasis on the Non-Enhancing Brain Tumor Compartment

T. Schaffer; A. Brawanski; S. Wein; A. M. Tomé; E. W. Lang

cs.CVFeb 25, 2026

Brain Tumor Segmentation with Special Emphasis on the Non-Enhancing Brain Tumor Compartment

Authors: T. Schaffer, A. Brawanski, S. Wein, A. M. Tomé, E. W. Lang

Organizations: CIML Group, Biophysics, University of Regensburg, 93040 Regensburg, Germany · Department of Neurosurgery , University Hospital Regensburg, 93042 Regensburg, Germany · Department of Biomedical Imaging, University Hospital Regensburg, 93042 Regensburg, Germany · DETI, IEETA, Universidade de Aveiro, 3810-195 Aveiro, Portugal

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.

Figures & tables

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

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

Panel of Figure 3 — Figure 3: Records from the BraTS 2018 dataset [ 9 ] named CBICA-ANI-1, CBICA-AXO-1, CBICA-AQQ-1, TCIA03-419 (top to bottom). Left: Flair (CBICA-ANI-1, CBICA-AXO-1) or T1C (CBICA-AQQ-1, TCIA03-419) slice. Center: original BraTS 2018 segmentation mask (ED yellow, NCR+NET red, ET blue). Right: NCR (red) and NET (green) separated and filtered.

Model	Parameters	Input size	Training speed
Upscaling PAU-Net 5-level	13.15M	$80\times 160\times 128(1.64M)$	696 ms/step
Upscaling PAU-Net 4-level	2.87M	$96\times 192\times 160(2.95M)$	1096 ms/step

Table 1: Model comparison regarding number of parameters, input size and training speed (measured on a Tesla A100-SXM4-40GB GPU)

Panel of Figure 4 — Figure 4: Sample record from BraTS 2021 dataset [ 8 ] . Top row from left to right: T1, T2, T1C, Flair; Bottom row left: BraTS 2021 tumor segmentation data, right: tumor segmentation data extended with NET region (yellow)

	BraTS 2018 train		BraTS 2018 test		BraTS 2021 train		BraTS 2021 test
	5-level	4-level	5-level	4-level	5-level	4-level	5-level	4-level
ET	0.7162	0.7179	0.7842	0.7802	0.8524	0.8566	0.8357	0.8464
TC	0.8803	0.8646	0.8125	0.8230	0.9140	0.9121	0.8798	0.8841
WT	0.9117	0.9010	0.8913	0.8912	0.9282	0.9232	0.9157	0.9186
Mean	0.8361	0.8278	0.8293	0.8315	0.8982	0.8973	0.8771	0.8830

Table 2: Upscaling PAU-Net results, Dice-Sørensen coefficients for Enhancing Tumor (ET), Tumor Core (TC) and Whole Tumor (WT)

Panel of Figure 5 — Figure 5: Results for training the PAU-Net on BraTS 2018 subset with extracted NET segments. Training was performed with 100 epochs, 1044ms/step, learning rate $10^{-4}$ with 5% decay per epoch, training data size 68, batch size 1, GPU Tesla T4. Left: final DSCs, Right: Learning curves

Figure 6: Gamma distribution fit to volume histograms

Eval. Segment	BraTS 2018/21 Data	BraTS 2018/21 Label	Description
ET	ET	4	Enhancing Tumor
TC	ET + NCR	4,1	Tumor Core
WT	ET + NCR + NET + ED	4,1,3,2	Whole Tumor
NET	NET	3	Non-Enhancing Tumor
TCN	ET + NCR + NET	4,1,3	Tumor Core + Non-Enhancing Tumor

Table 3: Unified BraTS 2018/21 evaluation segments as compositions of data labels. NET is added as a new evaluation segment as well as TCN, which extends the traditional TC labeled segment by the NET segment.

	SegResNetVAE	TSC U-Net	nnU-Net	SegResNet	nnU-Net	PAU-Net
	2018 [ 67 ]	2019 [ 38 ]	2020 [ 37 ]	2021 [ 80 ]	2021 [ 56 ]
NET						0.5320
TCN						0.8846
ET	0.8145	0.8021	0.7945	0.8600	0.8823	0.8194
TC	0.8596	0.8647	0.8524	0.8868	0.9235	0.8979
WT	0.9042	0.9094	0.9119	0.9265	0.9383	0.9151

Table 4: Dice-Sørensen Coefficient test results obtained with unified 4-label BraTS 2018/21 data. Our mean values were computed from ET, TC and WT, not taking into account NET and TCN.

Explore similar work

Lightweight Vision Transformer-Based U-Net for Brain Tumor Segmentation from MRI
Sep 24, 2026Sheekar Banerjee, Md. Srabon Chowdhury, Md. Mahbub Hasan Akash +1Hybrid CNN-Transformer Architectures Vision Transformer
Department of CSE, IUBAT, Dhaka, Bangladesh
A Novel Global Context-aware Deep Neural Network for Enhanced Brain Tumor Segmentation using Magnetic Resonance Images
May 28, 2026Sourjya Mukherjee, Ananya Bhattacharjee, R. MuruganU-Net Medical Image Segmentation
Biomedical Imaging Laboratory, Department of Electronics and Communication Engineering, National Institute of Technology Silchar, Silchar, Assam, India, 788010
DALight-3D: A Lightweight 3D U-Net for Brain Tumor Segmentation from Multi-Modal MRI
May 6, 2026Nand Kumar Mishra, Dhruv Mishra, Dr Manu Pratap SinghInteractive Medical Image Segmentation Image Segmentation
Department of Computer Science, Dr. Bhimrao Ambedkar University, Agra, India. · Department of Computer Science and Engineering, Shiv Nadar University, Greater Noida, India.

Sep 24, 2026J/K move · Enter open · S save

Lightweight Vision Transformer-Based U-Net for Brain Tumor Segmentation from MRI

Sheekar Banerjee, Md. Srabon Chowdhury, Md. Mahbub Hasan Akash, Ishtiak Al Mamoon

Department of CSE, IUBAT, Dhaka, Bangladesh

Hybrid CNN-Transformer Architectures Vision Transformer U-Net Medical Image Segmentation Brain Tumor Segmentation

Open paper AskPDF arXiv

Accurate brain tumor segmentation from Magnetic Resonance Imaging is essential for diagnosis, treatment planning, and surgical guidance. Although Convolutional Neural Networks, particularly UNet, have achieved significant success in medical image segmentation, they often struggle to capture the long-range spatial dependencies required to model tumors with irregular shapes and complex boundaries. This paper proposes a lightweight Vision Transformer UNet that combines the hierarchical feature extraction capability of UNet with the global context modeling of Vision Transformers. The proposed architecture incorporates a compact ViT bottleneck within a U-Net encoder-decoder framework, enabling effective learning of both local and global features while maintaining computational efficiency with only 2.6 million trainable parameters. The model was evaluated on the TCGA LGG MRI Segmentation dataset, achieving a mean Intersection over Union of 0.8100 and a Dice score of 0.8446, outperforming the baseline UNet by 3.75% and 3.15%, respectively. Extensive quantitative and qualitative analyses, including confusion matrix evaluation, precision recall curves, per-image performance distribution, and tumor size dependency analysis, demonstrate the effectiveness and robustness of the proposed method for brain tumor segmentation.

Abstract

Explore similar work

Lightweight Vision Transformer-Based U-Net for Brain Tumor Segmentation from MRI

A Novel Global Context-aware Deep Neural Network for Enhanced Brain Tumor Segmentation using Magnetic Resonance Images

DALight-3D: A Lightweight 3D U-Net for Brain Tumor Segmentation from Multi-Modal MRI