A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Stroke Research

Background:

• Accurate lesion segmentation in T1-weighted MRIs is crucial for stroke research but current automated methods lack reliability.
• Manual segmentation is the gold standard but is time-consuming and subjective.
• Previous datasets like ATLAS v1.2 have limitations, leading to suboptimal algorithm development.

Purpose of the Study:

• To introduce ATLAS v2.0, a significantly larger dataset of T1-weighted MRIs and manually segmented lesion masks for stroke research.
• To facilitate the development of more accurate and robust automated lesion segmentation algorithms.
• To provide a platform for unbiased evaluation of segmentation algorithms through hidden test and generalizability datasets.

Main Methods:

• The ATLAS v2.0 dataset comprises 1271 T1-weighted MRIs with manual lesion masks.
• The dataset is divided into training (n=655), test (n=300, hidden masks), and generalizability (n=316, hidden MRIs and masks) sets.
• The hidden datasets enable unbiased performance evaluation via segmentation challenges.

Main Results:

• The larger dataset size (N=1271) is expected to drive the development of more robust segmentation algorithms.
• The structured dataset with hidden components facilitates rigorous and unbiased algorithm validation.
• Anticipated outcome is improved automated lesion segmentation accuracy for T1w MRIs.

Conclusions:

• ATLAS v2.0 represents a significant advancement over previous datasets for stroke lesion segmentation research.
• This enhanced dataset is expected to foster the creation of superior algorithms, thereby advancing large-scale stroke research.
• The availability of ATLAS v2.0 will aid in overcoming current limitations in automated stroke lesion segmentation.