Deep Learning for Brain Tumor Segmentation (2026): U-Net, DUCK-Net

Six U-Net-Based Models, One Dataset, One Training Setup

10 entries

A LinkedIn post by Trần Quang Huyên (26 October 2026) describes a paper, "Comparative Analysis of Deep Learning Architectures for Brain Tumor Segmentation", published in the proceedings of ATC 2025, the International Conference on Advanced Technologies for Communications in Hanoi, and available on IEEE Xplore. The authors say each architecture had been evaluated before but rarely side by side under the same conditions, which makes published numbers hard to compare. They trained U-Net, ResUNet, DeepResUNet, two Recurrent U-Net variants (DRU and SRU) and DUCK-Net with one unified configuration on the public Figshare brain tumor dataset of 3,064 annotated MRI slices, and scored them with Dice and IoU. According to the post, DUCK-Net was the clear winner with a test Dice of 0.8575 and IoU of 0.7561; U-Net, ResUNet and DeepResUNet finished almost level at around 0.805 Dice; DRU and SRU used fewer than half the parameters of U-Net (13.9M and 12.7M against 31.3M) while reaching 0.7954 and 0.7825 Dice; and most models still struggled with small tumors and poorly defined boundaries. The post says the code and trained weights are public. These figures are the authors' own reported results, have not been reproduced here, and apply to that dataset and setup only. This page is educational research information, not medical advice, and segmentation models are not diagnostic tools. It lists the paper, code, datasets, metrics, related architectures and tools. Descriptions are paraphrased from each source and are not rankings.

01Six U-Net-based models, one MRI dataset (LinkedIn post by Trần Quang Huyên)The post summarizing the comparative study and its findings. 02Comparative Analysis of Deep Learning Architectures for Brain Tumor Segmentation (IEEE Xplore)The ATC 2025 conference paper linked from the post; the IEEE site may need a manual click. 03Comparative-Analysis-of-Deep-Learning-Architectures-for-Brain-Tumor-Segmentation (GitHub)Public code linked from the post for the study. 04HealthcareHealthcare hub, on this site. 05Healthcare AI AgentsAgents that turn clinical data into action, on this site. 06Deep LearningDeep learning hub, on this site. 07How Vision Transformers (ViT) WorkPatches, attention and classification, on this site. 08Image Data Analytics SolutionsImage analytics tools, on this site. 09Machine Learning ModelsMachine learning model directory, on this site. 10Vision-Language ModelsModels that combine images and text, on this site.

1. The Comparative Study

6 entries

The paper, code and the evaluation setup described in the post.

11Paper on IEEE XploreATC 2025 proceedings entry for the comparative analysis; the IEEE site may need a manual click. 12Study code and weights (GitHub)Repository linked from the post, which says code and trained weights are public. 13Figshare brain tumor datasetThe public dataset of annotated MRI slices used in the study, per the post; the Figshare site may need a manual click. 14BraTS challenge (CBICA)Benchmark challenge series for brain tumor segmentation. 15BraTS on SynapseProject page for the BraTS challenges. 16Identifying the best ML algorithms for brain tumor segmentation (arXiv)BraTS challenge analysis of segmentation, progression assessment and survival prediction.

2. The Architectures Compared

5 entries

U-Net and its residual, recurrent and DUCK variants.

17U-Net (arXiv)The original convolutional network for biomedical image segmentation. 18Road Extraction by Deep Residual U-Net (arXiv)Paper introducing a residual U-Net, the basis for ResUNet-style models. 19R2U-Net (arXiv)Recurrent residual convolutional neural network for medical image segmentation, related to recurrent U-Net variants. 20Using DUCK-Net for Polyp Image Segmentation (arXiv)Paper introducing DUCK-Net, applied there to polyp segmentation. 21Pytorch-UNetA PyTorch implementation of U-Net for image segmentation.

3. Metrics and Evaluation

3 entries

Dice and IoU score overlap between predicted and annotated tumor regions; small tumors and unclear boundaries remain hard.

22MONAI metrics and lossesMONAI is an AI toolkit for healthcare imaging with Dice, IoU and related metrics; GitHub lists 1.6.1 (27 September 2026). 23MONAI repositorySource and documentation for the MONAI toolkit. 24segmentation_models.pytorchLibrary of semantic segmentation models, encoders and losses; GitHub lists v0.6.0 (26 October 2026).

4. Beyond U-Net: Related Approaches

8 entries

Other architectures and frameworks widely used for medical image segmentation.

25nnU-Net (arXiv)Self-adapting framework for U-Net-based medical image segmentation. 26nnU-Net repositorySource for nnU-Net from the DKFZ. 27Attention U-Net (arXiv)Attention gates that help a U-Net focus on relevant regions. 28TransUNet (arXiv)Combines Transformers with U-Net for medical image segmentation. 29Swin UNETR (arXiv)Swin Transformers for semantic segmentation of brain tumors in MRI images. 30Segment Anything in Medical Images (arXiv)MedSAM, adapting a promptable segmentation model to medical images. 31MedSAM repositoryCode for Segment Anything in Medical Images. 32How Vision Transformers (ViT) WorkBackground on the Transformer backbones behind TransUNet and Swin UNETR, on this site.

Frequently Asked Questions

What did the brain tumor segmentation study compare?

Six U-Net-based models: U-Net, ResUNet, DeepResUNet, two Recurrent U-Net variants (DRU and SRU) and DUCK-Net, all trained with one unified configuration on the public Figshare brain tumor dataset of 3,064 annotated MRI slices, as the LinkedIn post this page follows reports.

Which model performed best?

According to the post, DUCK-Net had the best test results, with a Dice of 0.8575 and an IoU of 0.7561. U-Net, ResUNet and DeepResUNet were almost level at around 0.805 Dice. These are the authors' own results on that dataset and setup.

Are smaller models still useful?

The post says the two Recurrent U-Net variants use fewer than half the parameters of U-Net (13.9M and 12.7M against 31.3M) and still reached 0.7954 and 0.7825 Dice, a trade-off when compute and memory are limited.

What are Dice and IoU?

Overlap metrics that compare a predicted segmentation with the annotated region. Higher values mean greater overlap. The study scored models with both.

What are the remaining challenges?

The post says most models still struggled with small tumors and poorly defined boundaries.

Can I reuse the code and weights?

The post says the code and trained weights are public. Check the repository and its license, and validate any model on your own data before relying on it.

Can these models be used for diagnosis?

No. This page is educational research information, not medical advice. Segmentation research models are not diagnostic tools, and clinical use requires validation and regulatory approval.

How current is this page?

Links and release facts were checked on 1 November 2026. LinkedIn and IEEE may ask visitors to sign in or click manually. Descriptions follow each source's own pages and are not rankings or guarantees.