Source-linked AI summary
BRISC: Annotated Dataset for Brain Tumor Segmentation and Classification
Amirreza Fateh, Yasin Rezvani, Sara Moayedi, Sadjad Rezvani, Fatemeh Fateh, Mansoor Fateh, Vahid Abolghasemi
TL;DR
Brain-tumor MRI segmentation and classification are constrained by limited high-quality, diverse, expertly labeled datasets. BRISC constructs and releases a 6,000-image, multi-class, multi-plane dataset with expert-verified annotations and benchmarks, while remaining limited by incomplete acquisition metadata and uncertain subject-level split independence.
Problem
High-quality labeled MRI datasets for automated brain-tumor detection, segmentation, and classification are scarce, while existing datasets may rely on standardized preprocessing or underrepresent tumor types.
Method
BRISC curates 6,000 contrast-enhanced T1-weighted MRI scans from public datasets, adds expert annotations, and evaluates segmentation and classification with standard deep learning models.
Results
Benchmark results establish strong baselines, including 80.6% weighted mIoU for SaberNet and 0.9920 ± 0.0000 weighted F1-score for EfficientNetB0.
Takeaways & Limitations
BRISC provides a public foundation for benchmarking and developing brain-tumor segmentation and classification models across tumor categories and imaging planes.
Takeaways & Limitations
The source data lack comprehensive acquisition metadata, complete subject-level independence across splits cannot be guaranteed, and models may experience domain shift on other imaging settings.
Abstract
from arXiv · showhide
Accurate segmentation and classification of brain tumors from Magnetic Resonance Imaging (MRI) remain key challenges in medical image analysis, primarily due to the lack of high-quality, balanced, and diverse datasets with expert annotations. In this work, we address this gap by introducing BRISC, a dataset designed for brain tumor segmentation and classification tasks, featuring high-resolution segmentation masks. The dataset comprises 6,000 contrast-enhanced T1-weighted MRI scans, which were collated from multiple public datasets that lacked segmentation labels. Our primary contribution is the subsequent expert annotation of these images, performed by certified radiologists and physicians. It includes three major tumor types, namely glioma, meningioma, and pituitary, as well as non-tumorous cases. Each sample includes high-resolution labels and is categorized across axial, sagittal, and coronal imaging planes to facilitate robust model development and cross-view generalization. To demonstrate the utility of the dataset, we provide benchmark results for both tasks using standard deep learning models. The BRISC dataset is made publicly available. datasetlink: https://www.kaggle.com/datasets/briscdataset/brisc2025/
Background & Summary
BRISC addresses the scarcity of high-quality labeled MRI datasets for automated brain-tumor detection, segmentation, and classification. It provides a balanced, expert-annotated, multi-plane dataset spanning four categories.
- Automated brain-tumor detection and segmentation remain challenging because high-quality labeled datasets are scarce and tumor appearances vary across patients.
- Existing datasets have limitations including reliance on standardized preprocessing and limited representation of tumor types beyond glioma.
- BRISC contains 6,000 contrast-enhanced T1-weighted MRI scans covering glioma, meningioma, pituitary, and non-tumorous cases across axial, coronal, and sagittal planes.
- The dataset emphasizes expert annotation, balanced distributions, consistent quality, and multi-institutional diversity for segmentation and classification model development.
Methods
The dataset was curated from public MRI collections through sequence selection, quality control, deduplication, standardization, expert review, and annotation. Its intended use is algorithm benchmarking rather than direct clinical diagnosis.
- Dataset design: BRISC contains balanced labels for glioma, meningioma, pituitary tumors, and non-tumorous cases, addressing class imbalance and annotation inconsistencies in existing datasets.
- Dataset design: The 6,000 images are divided into 5,000 training and 1,000 testing images, with class and MRI-plane distributions designed to be nearly uniform.
- Curation: Curation retained T1-weighted scans and applied label verification, artefact screening, deduplication before splitting, and spatial standardization.
- Curation: Single-image sequences were excluded because annotators found tumor identification difficult in those cases.
- Curation: Patient-level independence across splits cannot be guaranteed because the source data lacked patient, sequence, and slice identifiers, although likely same-subject cases were manually separated when possible.
- Annotation: Annotations were iteratively refined and verified by trained annotators under certified physician and radiologist supervision.
- Annotation: A representative quality assessment found a mean Dice coefficient of 0.924 between initial and expert-verified masks, with approximately 4.8% of images requiring correction.
- Limitations and intended use: BRISC lacks comprehensive acquisition metadata, cannot be fully harmonized across hardware and protocols, and is intended for research benchmarking rather than direct clinical diagnostic use.
Data Records
BRISC is publicly released with images, masks, manifests, metadata, checksums, and a documented directory structure. Records use standardized filenames and link segmentation masks to their corresponding images.
- Release contents: The public release includes dataset images, a CSV manifest, JSON metadata, file checksums, and per-file metadata.
- Release contents: The release follows the directory structure shown in Figure 1.
- File organization: Filenames encode the dataset name, split, index, tumor code, plane code, sequence, and file extension.
- File organization: Segmentation image-mask pairs share a basename, with images stored as JPEG files and masks as PNG files.
- Metadata: Manifest rows record paths, task, split, tumor and plane labels, sequence, dimensions, file size, and SHA-256 checksum; mask rows link to their matched images.
- Provenance: BRISC was derived from a collection aggregating the Cheng, SARTAJ, and Br35H brain-tumor datasets.
Data Overview
BRISC provides separate classification and segmentation records using 2D single-slice T1-weighted brain MRI images with pixel-wise masks. The classification task contains 6,000 images, while segmentation contains 4,793 paired images and masks.
- Data format: BRISC stores 2D single-slice T1-weighted brain MRI images as JPEG files with corresponding pixel-wise segmentation masks as PNG files.
- Classification task: The classification task contains 6,000 JPEG images across glioma, meningioma, pituitary tumor, and no-tumor classes, split into 5,000 training and 1,000 test images.
- Segmentation task: The segmentation task contains 4,793 image files with exact paired masks for pixel-wise tumor annotation.
Technical Validation
BRISC is validated through baseline evaluations for segmentation and classification, using task-specific metrics and diverse standard models. The benchmarks show strong segmentation by transformer-based models and high classification performance from EfficientNetB0, while performance varies substantially across architectures.
- Validation design: BRISC validation evaluates segmentation and classification using baseline models and reports task-specific performance metrics.The study uses mean and weighted mIoU for segmentation and per-class, macro, and weighted classification metrics.
- Segmentation metrics: IoU measures overlap between predicted and ground-truth tumor regions, with a small constant added for numerical stability.The formulation is intended to assess pixel-wise segmentation quality, including object boundaries.
- Segmentation metrics: Weighted mIoU accounts for the proportion of samples in each tumor type when summarizing segmentation performance.The metric is reported alongside tumor-specific mIoU for glioma, meningioma, and pituitary cases.
- Segmentation results: 80.6% weighted mIoU was achieved by SaberNet, compared with 75.7% for UNet in the segmentation benchmarks.The comparison spans convolutional, attention-enhanced, and transformer-based architectures.
- Classification results: 0.9920 ± 0.0000 weighted average F1-score and 0.9920 ± 0.0000 accuracy were reported for EfficientNetB0 in classification.EfficientNetB0 also achieved 1.0000 ± 0.0000 recall for the non-tumorous class.
- Classification results: Classification performance varied sharply across architectures, from EfficientNetB0’s strong results to Xception’s 0.1780 ± 0.0658 weighted F1-score.Xception failed entirely on meningioma and pituitary, while DenseNet variants showed unstable performance and MobileNetV2 had very low meningioma recall.