IntrA: 3D Intracranial Aneurysm Dataset for Deep Learning (2003.02920v2)

Abstract: Medicine is an important application area for deep learning models. Research in this field is a combination of medical expertise and data science knowledge. In this paper, instead of 2D medical images, we introduce an open-access 3D intracranial aneurysm dataset, IntrA, that makes the application of points-based and mesh-based classification and segmentation models available. Our dataset can be used to diagnose intracranial aneurysms and to extract the neck for a clipping operation in medicine and other areas of deep learning, such as normal estimation and surface reconstruction. We provide a large-scale benchmark of classification and part segmentation by testing state-of-the-art networks. We also discuss the performance of each method and demonstrate the challenges of our dataset. The published dataset can be accessed here: https://github.com/intra3d2019/IntrA.

• The paper introduces IntrA, a comprehensive 3D aneurysm dataset enabling enhanced classification and segmentation in deep learning.
• It details the reconstruction of 3D models from 103 patient MRA images, producing 1909 vessel segments and 116 expert-annotated aneurysm regions.
• Benchmark tests achieved over 88% detection accuracy and above 80% segmentation scores, underscoring the dataset's potential to improve neurosurgical planning.

IntrA: A Novel 3D Intracranial Aneurysm Dataset for Advancing Deep Learning Applications in Medical Science

The paper addresses the significant gap in medical datasets suitable for 3D deep learning applications by introducing IntrA, an open-access 3D dataset focusing on intracranial aneurysms. Unlike the prevalent 2D medical imagery, this dataset provides comprehensive 3D intracranial aneurysm models annotated for classification and segmentation tasks in deep learning. IntrA represents a critical asset for advancing 3D mesh and point-based deep learning applications, particularly in the medical field where diagnostic accuracy can be greatly enhanced.

Dataset Overview

IntrA comprises three categories of data: complete models of brain vessels, automatically generated vessel segments, and manually annotated aneurysm segments by medical experts. A total of 103 patient-derived 3D models were reconstructed from 2D Magnetic Resonance Angiography (MRA) images, resulting in 1909 automatically generated segments and 116 expert-annotated aneurysm segments. This extensive dataset enables researchers to explore more complex analyses, such as the segmentation of aneurysm necks, which are crucial for planning surgical interventions.

Methodological Contributions

The paper presents several methodological contributions:

• An accessible dataset featuring 3D aneurysm segments differentiated by segmentation annotations and complete models of scanned blood vessels.
• Development of interactive tools that facilitate the generation and annotation of 3D aneurysm models.
• Extensive benchmarking of state-of-the-art 3D deep learning methods for classifying and segmenting intracranial aneurysms.

Benchmark Results

The authors tested various leading-edge 3D deep learning techniques on the IntrA dataset to establish benchmarks. These include PointNet, PointNet++, PointCNN, SpiderCNN, SO-Net, and others, focusing on both classification and segmentation tasks.

Classification Results:

• The most effective models achieved an aneurysm detection accuracy exceeding 88%. Notably, PointNet++ with 1024 input samples yielded the highest detection rate, highlighting the importance of robust 3D point cloud methodologies.

Segmentation Results:

• Segmentation performance varied, with SO-Net and PointConv scoring the highest in segmenting aneurysm regions, achieving Jaccard Indices (IoU) and Dice Coefficient (DSC) values above 80%.
• The paper indicated that incorporating geodesic information significantly boosts the segmentation accuracy, demonstrating more refined surface differentiation capabilities of point cloud networks.

Implications and Future Directions

The availability of the IntrA dataset opens new avenues for developing and testing deep learning models suited for 3D medical images. The paper emphasizes that current methods, while impressive on traditional datasets, require adaptation to handle the complex, non-Euclidean manifold structures presented by 3D medical geometries effectively.

From a practical standpoint, better aneurysm segmentation holds potential for enhancing clinical decision-making in neurosurgical planning. Theoretically, the dataset encourages further investigation into structure-based learning paradigms capable of leveraging the geodesic information of medical geometries.

Future research could focus on expanding the dataset and integrating synthetic data for augmentation to overcome the limitations imposed by the current dataset size. The application of novel network architectures designed specifically for medical datasets could provide real-world clinical tools, further bridging the gap between computational research and medical practice.

In conclusion, IntrA represents a considerable step forward in the synthesis of medical science and computational methodologies, providing a benchmark for 3D segmentation and classification tasks that have traditionally relied on 2D data representations.