Source-linked AI summary
Using U-Net Network for Efficient Brain Tumor Segmentation in MRI Images
Jason Walsh, Alice Othmani, Mayank Jain, Soumyabrata Dev
TL;DR
Brain tumor segmentation from MRI is important because manual identification is onerous and difficult. This paper proposes a lightweight U-Net that uses two-dimensional views of MRI volumes, and it reports strong benchmark performance while avoiding aggressive augmentation.
Problem
Manual brain tumor identification is onerous and difficult, while existing deep-learning methods may require large amounts of annotated data or aggressive augmentation.
Method
The paper proposes a lightweight U-Net that converts three-dimensional MRI brain scans into two-dimensional images and uses the three perspective planes for segmentation.
Results
84% mean IoU was achieved when training on the entire dataset, and the proposed method outperformed all four benchmark methods across each perspective plane and the full dataset.
Takeaways & Limitations
The proposed architecture provides accurate brain tumor segmentation without aggressive data augmentation.
Takeaways & Limitations
The authors state that more studies are needed to further improve the performance of the proposed learning architecture.
Abstract
from arXiv · showhide
Magnetic Resonance Imaging (MRI) is the most commonly used non-intrusive technique for medical image acquisition. Brain tumor segmentation is the process of algorithmically identifying tumors in brain MRI scans. While many approaches have been proposed in the literature for brain tumor segmentation, this paper proposes a lightweight implementation of U-Net. Apart from providing real-time segmentation of MRI scans, the proposed architecture does not need large amount of data to train the proposed lightweight U-Net. Moreover, no additional data augmentation step is required. The lightweight U-Net shows very promising results on BITE dataset and it achieves a mean intersection-over-union (IoU) of 89% while outperforming the standard benchmark algorithms. Additionally, this work demonstrates an effective use of the three perspective planes, instead of the original three-dimensional volumetric images, for simplified brain tumor segmentation.
1. Introduction
Brain tumor segmentation uses MRI to identify tumor size and location, but deep-learning approaches may require extensive annotation or augmentation. This paper proposes a lightweight U-Net using three two-dimensional perspective planes and benchmarks it against existing algorithms.
- MRI is widely used for brain tumor diagnosis because it is a non-invasive imaging technology for the complex and sensitive brain.
- Brain tumor segmentation aims to automatically identify tumor size and location from MRI scans.
- Deep-learning segmentation methods have achieved success but may require large amounts of annotated data or aggressive data augmentation.
- The proposed lightweight U-Net uses three perspective planes instead of original three-dimensional volumetric images for simplified brain tumor segmentation.
- The study benchmarks the proposed model against several widely used segmentation algorithms.
2. Brain Tumor Segmentation Techniques
Brain tumor segmentation techniques range from thresholding, classifiers, and clustering to deep-learning architectures operating on two- or three-dimensional images. The reviewed methods use varied preprocessing, feature-learning, and segmentation strategies.
- 2.1. Thresholding: Thresholding methods convert images to binary form, with variants using Otsu’s or global thresholding to extract regions of interest.
- 2.2. Classifiers: Classifiers learn from annotated images, while k-nearest neighbour segmentation merges pixels with similar computed features.
- 2.2. Classifiers: A prior HMM approach applied to two-dimensional MRI scans from the BITE dataset produced probabilities for cancerous and non-cancerous categories.
- 2.3. Clustering: Clustering methods such as k-means and fuzzy c-means identify patterns without a target labelled dataset and can segment areas of interest from backgrounds.
- 2.4. Deep Learning: Deep-learning architectures such as U-Net and V-Net segment regions of interest from two- or three-dimensional images.
- 2.4. Deep Learning: LinkNet uses a lightweight architecture to reduce parameters and enable real-time segmentation.
3. Magnetic Resonance Image (MRI) data
The study reformats three-dimensional MRI data into two-dimensional slices from coronal, sagittal, and transversal planes for brain-tumor segmentation. It uses the BITE dataset and addresses inconsistent or missing ground-truth information through mask processing and dataset selection.
- Dataset: BITE provides pre-operative, post-operative, and ground-truth images for 14 patients, with the analysis focused on group three.The scans are T1-weighted MRI accompanied by B-mode ultrasound images and stored in MINC format.
- Dataset limitations: The analysis excludes patient 13 because no pre-operative scan exists, preventing manual extraction of its pre-operative ROI mask.
- Dataset limitations: The ground-truth masks have patient-dependent resolutions that do not match the pre- or post-operative images, creating a dataset constraint for segmentation.
- Preprocessing: Because U-Net requires matching image and mask dimensions, the binary masks are resampled to align with the corresponding pre-operative scans.Resampling adjusts mask height, width, and slice count so the data can be used for segmentation.
- Visualization: Overlaying a resampled tumor mask on the pre-operative MRI visualizes the tumor’s size, shape, and location.
4. Methodology
The methodology adapts U-Net for lightweight two-dimensional MRI segmentation by extracting perspective-plane slices and using a compact convolutional encoder–decoder. Training uses binary cross-entropy, Adam optimization, monitoring callbacks, and experiments varying epochs and filter counts.
- U-Net architecture: U-Net uses an encoder to capture image context and a symmetric decoder with transposed convolutions for precise localization.
- Data preparation: The dataset is converted from three-dimensional scans into four two-dimensional datasets: one for each perspective plane and one containing all available images.The slice extractor saves MRI slices as PNG files.
- Method: The proposed implementation uses four convolutional blocks with 3 × 3 kernels, ReLU activations, batch normalization, and 2 × 2 max pooling.Filter sizes increase in steps of 16 across convolutional blocks.
- Training & Optimization: Binary cross-entropy trains the binary tumor-region predictor, whose sigmoid output is compared with target values of 1 or 0.The loss is defined over the output size N.
- Method: The network is designed to be lightweight and swift so it can reproduce prediction images from the selected input size.The input dimensions were fixed at 128 × 128 because image sizes vary across perspective planes.
- Training & Optimization: Adam optimization uses a learning rate of 0.0001, while early stopping, checkpointing, and CSV-logged precision, recall, and IoU monitor training.
5. Results
The proposed U-Net achieved strong segmentation with few parameters and short training, while increasing filters improved IoU and convergence across MRI perspective planes. It outperformed benchmark methods and required no data augmentation.
- 50 epochs improved segmentation performance across all three perspective planes, while extreme epoch increases produced only slight additional accuracy.
- The network produced accurate segmentation with few trainable parameters and a small epoch size, without requiring data augmentation.
- Increasing filter values raised stepwise IoU to 71%, significantly above the standard network, and improved segmentation performance.
- The first small-parameter U-Net achieved 70%–80% average mean IoU across the three perspective planes and 84% on the full dataset.
- 89% mean IoU was reported for the lightweight U-Net on the BITE dataset.
- The proposed model outperformed each of four benchmark algorithms across all four evaluation metrics and across the perspective planes and full dataset.
6. Conclusion
The study presents a lightweight U-Net variant that segments brain-tumor anomalies from two-dimensional MRI images derived from three-dimensional scans. On the BITE dataset, it outperforms standard benchmark algorithms while requiring neither aggressive data augmentation nor large datasets.
- The method converts BITE’s three-dimensional MRI scans into two-dimensional images representing the three perspective planes for segmentation.
- The lightweight U-Net accurately segments anomalies from two-dimensional images without requiring aggressive data augmentation.
- 84% mean IoU was achieved when the lightweight U-Net was trained on the entire dataset, outperforming the standard benchmark algorithms.
- The architecture is presented as a suitable approach for segmenting anomalies in small datasets containing fewer than one thousand images.
- The approach could support physicians by providing a second evaluator for patients’ MR images in medical settings.
- The authors identify false-negative and false-positive ratios as important targets for further performance improvement and propose benchmarking against original U-Net and other networks.