Source-linked AI summary
Deep Learning Based Brain Tumor Segmentation: A Survey
Zhihua Liu, Lei Tong, Zheheng Jiang, Long Chen, Feixiang Zhou, Qianni Zhang, Xiangrong Zhang, Yaochu Jin, Huiyu Zhou
TL;DR
Brain tumor segmentation is a challenging medical-image analysis problem requiring accurate delineation of tumor regions. This survey categorizes and compares deep-learning methods across network architectures, imbalanced conditions, and multimodality, drawing on more than one hundred studies. It identifies recurring strategies and trade-offs in architecture design, imbalance handling, and multimodal feature fusion, while outlining open challenges and future directions.
Problem
Accurate brain tumor segmentation is difficult because tumors exhibit location and morphological uncertainty, while single-task approaches may overlook information from related tasks.
Method
The survey structurally categorizes and technically reviews deep-learning segmentation methods using an architecture, imbalance, and multimodality taxonomy.
Results
Multimodal fusion can improve feature expressiveness and generalization, while modality ranking and pairing help select useful modalities but may require additional computation or parameters.
Takeaways & Limitations
The taxonomy and comparisons organize technical similarities and differences and help identify open challenges and future research directions.
Takeaways & Limitations
Network and module design relies heavily on human experience and generally does not incorporate brain-tumor domain knowledge such as degree or physically inspired morphological information.
Abstract
from arXiv · showhide
Brain tumor segmentation is one of the most challenging problems in medical image analysis. The goal of brain tumor segmentation is to generate accurate delineation of brain tumor regions. In recent years, deep learning methods have shown promising performance in solving various computer vision problems, such as image classification, object detection and semantic segmentation. A number of deep learning based methods have been applied to brain tumor segmentation and achieved promising results. Considering the remarkable breakthroughs made by state-of-the-art technologies, we use this survey to provide a comprehensive study of recently developed deep learning based brain tumor segmentation techniques. More than 100 scientific papers are selected and discussed in this survey, extensively covering technical aspects such as network architecture design, segmentation under imbalanced conditions, and multi-modality processes. We also provide insightful discussions for future development directions.
1. Introduction
Brain tumor segmentation remains a difficult medical-imaging task, while MRI-based delineation supports diagnosis and treatment. This survey reviews deep-learning methods through a technical taxonomy covering architectures, imbalance, and multimodality.
- Accurate brain tumor segmentation remains challenging because tumors vary in location and morphology, including size, shape, and sub-region structure.
- MRI segmentation assigns each voxel or pixel to tumor or predefined sub-region categories and returns a corresponding segmentation map.The task can use one or multiple image modalities, such as MRI sequences.
- Accurate glioma masks may support surgery planning, postoperative observation, and improved survival-rate outcomes.
- Prior surveys covered broad medical imaging, classic methods, or selected deep-learning applications, but lacked comprehensive technical analysis of deep-learning segmentation under imbalance and multimodality.
- The survey reviews more than one hundred studies from major journals, conferences, challenges, and preprint sources.
- Its taxonomy organizes methods by architecture design, imbalanced-condition strategies, and multimodality processing to compare technical similarities, differences, and open directions.
2. Background
Brain tumor segmentation remains difficult because tumors vary widely in location and morphology. This survey organizes deep learning methods into architectural design, imbalance handling, and multimodality categories, while reviewing their progress and contributions.
- Research Challenges: Tumors present location and morphological uncertainty, since gliomas may occur anywhere in the brain and vary substantially in shape and size.Edema and tumor sub-regions provide limited prior information about tumor shape.
- Progress in the Past Decades: Deep learning methods emerged after earlier traditional machine-learning approaches and later adopted customized CNNs, FCNs, and U-Net-based designs for tumor segmentation.The survey situates these developments within the field’s progression from hand-crafted features to deep neural networks.
- Progress in the Past Decades: The survey categorizes deep learning brain tumor segmentation methods into effective architectures, imbalanced-condition solutions, and multimodality approaches.This taxonomy is presented as a structured organization of the research field.
- Background: Its scope focuses on sub-region-level semantic segmentation rather than tasks such as tumor detection or single-label masks.The paper distinguishes its topic from brain-tissue segmentation, lesion segmentation, and tumor detection.
- Contributions of this survey: The survey reviews more than one hundred scientific papers and compares methods using public challenges and datasets.It includes journal articles, conference proceedings, annual challenges, and competition entries.
3. Designing Effective Segmentation Networks
Deep learning segmentation networks are designed to improve feature learning, localization, accuracy, and computational efficiency through specialized modules and architectures. The survey examines module innovations, path structures, encoder-decoder designs, and unresolved trade-offs including patch dependence and limited domain knowledge.
- Designing Effective Segmentation Networks: Effective network components either enlarge receptive fields and improve localization or reduce parameters and computational costs.The survey identifies semantic learning and efficiency as the two principal design directions.
- Designing Effective Architectures: Architectural development progresses from single-channel and fully connected designs toward multi-channel, fully convolutional, deep cascaded, and encoder-decoder networks.These changes aim to deepen feature learning and produce more precise segmentation.
- Modules for Higher Accuracy: Residual connections stabilize training in deeper networks by addressing gradient vanishing and explosion.They add a module’s input to its output while enabling deeper feature extraction.
- Modules for Higher Accuracy: Dilated convolutions and atrous spatial pyramid modules preserve spatial resolution while expanding receptive fields for subtle structures.These designs target localization and segmentation of small structures.
- Multi-Path Architecture: Single-path networks offer computational efficiency, whereas multi-path networks combine pathways to capture global and local features.Large-scale paths provide information about tumor location, size, and shape, while small-scale paths describe texture and boundaries.
- Multi-Path Architecture: Patch-based single- and multi-path networks are sensitive to patch size and quality, and fully connected layers may inadequately represent complex feature-to-label mappings.Small patches omit spatial information, whereas large patches require more computation.
- Summary: Existing architecture design relies heavily on human experience and often omits tumor-specific domain knowledge such as degree or physically inspired morphological information.The survey anticipates neural architecture search and domain-informed modeling as future directions.
4. Segmentation under Imbalanced Condition
The survey identifies data imbalance across tumor sub-regions, patient grades, and manual labels as a longstanding segmentation challenge. It reviews multi-network, multi-task, and customised-loss strategies, including cascades, ensembles, auxiliary tasks, and weighted losses.
- Data imbalance affects tumor sub-region pixels, HGG/LGG patient counts, and labels produced by different experts.The survey treats annotation differences as a special form of imbalance.
- Existing imbalance strategies are grouped into multi-network driven, multi-task driven, and custom loss function driven approaches.The taxonomy is organized around each strategy’s core component.
- Multi-network driven approaches: Cascaded networks pass upstream segmentation outputs downstream for coarse-to-fine refinement, but downstream quality depends heavily on upstream performance.The reviewed cascade designs may reduce imbalance interference by restricting later segmentation to upstream regions of interest.
- Multi-network driven approaches: Ensembling multiple homogeneous networks improves final segmentation results, with BraTS2018 reporting a 1% Dice gain over the best single network.The BraTS2019 winner ensembled 12 models and reported a smaller improvement of around 0.6–1%.
- Multi-task driven approaches: Multi-task learning shares low-level representations across related tasks and uses their joint feedback during back-propagation.The survey presents shared domain information and mutual restraint as advantages over single-task learning.
- Customised loss function driven approaches: Customised losses emphasize difficult or boundary samples; edge-pixel weighting improved segmentation Dice by 2–4%.Other reviewed losses address reconstruction, spatial information, focal weighting, class proportions, multi-scale supervision, and modality-specific objectives.
5. Utilising Multi Modality Information
The survey organizes multimodality processing around ranking modalities, pairing modality combinations, and fusing modality features. It also examines missing-modality settings, where generated replacements depend strongly on the quality of available data.
- Multimodality brain tumor segmentation uses MRI sequences such as T1, T1ce, T2, and Flair to exploit different tissue emphases.The survey distinguishes complete multimodality processing from clinical settings with missing modalities.
- Learning to Rank Modalities: Learning to rank models modality-task relevance so the network can focus on modalities most conducive to accurate segmentation.Existing approaches commonly learn modality-aware features while implicitly modeling the ranking.
- Learning to Pair Modalities: Learning to pair selects modality combinations for segmentation, but exhaustive pairing can require substantial computing resources.Reviewed methods model modality–modality relationships by evaluating combinations of modalities.
- Learning to Fuse Modalities: Learning to fuse combines features from different modalities, using addition or concatenation without extra parameters or attention modules with higher cost.Attention can highlight useful and suppress redundant features, while introducing parameters and computational expense.
- Dealing with Missing Modalities: Missing-modality methods generate absent modalities from available ones or model possible missing scenarios, and multimodality strongly influences accurate segmentation.The survey describes generating Flair from T1 and notes that generator quality depends heavily on the existing modality data.
6. Conclusion
The survey presents deep learning-based brain tumor segmentation as a challenging task and comprehensively organizes its methods and evaluation considerations. It covers method advantages, disadvantages, design motivations, and performance evaluation.
- The survey comprehensively investigates deep learning-based brain tumor segmentation methods and structurally categorizes their motivations, advantages, disadvantages, and evaluation.