BRISC: Annotated Dataset for Brain Tumor Segmentation and Classification

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/

💡 Research Summary

The paper introduces BRISC (Brain tumor Image Segmentation and Classification), a new publicly‑available dataset designed to address the persistent limitations of existing brain‑tumor imaging collections such as BraTS and Cheng. BRISC comprises 6,000 contrast‑enhanced T1‑weighted MRI scans sourced from multiple public repositories, all of which have been re‑annotated by certified radiologists and physicians. The dataset is balanced across four diagnostic categories—glioma (1,401 images), meningioma (1,635), pituitary tumor (1,757), and non‑tumorous cases (1,207)—and split into 5,000 training and 1,000 test images. Moreover, each image is labeled with its anatomical plane (axial, coronal, sagittal), and the distribution across planes is deliberately uniform (approximately 1,550 axial, 1,660 coronal, 1,790 sagittal in training; 387, 316, 297 respectively in testing).

Data curation involved several rigorous steps: (1) sequence harmonization to retain only T1‑weighted contrast‑enhanced scans, (2) thorough label and mask verification by a radiologist‑physician pair, (3) exclusion of corrupted or artifact‑laden files, (4) de‑duplication of exact and near‑duplicate slices, (5) spatial standardization (resizing, margin adjustment), and (6) manual cross‑checking to minimize patient overlap between splits. Although complete patient‑level independence cannot be guaranteed due to missing subject identifiers, the authors employed visual similarity checks and metadata cross‑reference to reduce leakage.

Annotation was performed using the AnyLabeling tool, with iterative refinement and consensus reviews. The resulting segmentation masks achieve a mean Dice coefficient of 0.924 when compared to initial drafts, and only 4.8 % of images required correction, indicating high annotation fidelity. The dataset is organized into two tasks: a classification task (6,000 JPEG images with class labels) and a segmentation task (4,793 image‑mask pairs, PNG masks). Detailed metadata files (manifest.csv and manifest.json) provide paths, class codes, plane codes, sequence information, image dimensions, and SHA‑256 checksums, facilitating reproducibility and easy integration into pipelines.

To demonstrate utility, the authors benchmarked a suite of state‑of‑the‑art models. For segmentation, twelve architectures—including classic CNNs (UNet, UNet++, DeepLabV3+, PAN) and recent transformer‑based networks (SABERNet, ABANet)—were evaluated using mean Intersection‑over‑Union (mIoU) across the three tumor types. CNN‑based models achieved mIoU scores between 0.78 and 0.84, while transformer‑enhanced models reached up to 0.86, confirming that BRISC can support both conventional and cutting‑edge approaches. For classification, ResNet‑50, EfficientNet‑B3, and ConvNeXt were trained, yielding an overall accuracy of 93.2 % and a weighted average F1‑score of 0.91, further underscoring the dataset’s quality.

The authors acknowledge key limitations: the exclusive focus on contrast‑enhanced T1‑weighted images restricts applicability to other sequences (e.g., T2, FLAIR) and may cause domain shift when models are transferred to heterogeneous clinical data. Additionally, scanner type, field strength, and acquisition parameters are absent, preventing fine‑grained harmonization across hardware. Consequently, BRISC is positioned primarily as a research benchmark rather than a clinically validated tool; users are cautioned to perform additional validation and potential fine‑tuning before clinical deployment.

In summary, BRISC fills a critical gap by providing a large, balanced, multi‑plane, expertly annotated brain‑tumor dataset. Its comprehensive documentation, open‑source availability on Kaggle, and baseline performance metrics make it an invaluable resource for developing robust segmentation and classification algorithms, exploring domain adaptation, and advancing multi‑view deep learning strategies in neuro‑oncology imaging.