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

Abstract

Accurate lesion segmentation is critical in stroke rehabilitation research for the quantification of lesion burden and accurate image processing. Current automated lesion segmentation methods for T1-weighted (T1w) MRIs, commonly used in stroke research, lack accuracy and reliability. Manual segmentation remains the gold standard, but it is time-consuming, subjective, and requires neuroanatomical expertise. We previously released an open-source dataset of stroke T1w MRIs and manually-segmented lesion masks (ATLAS v1.2, N = 304) to encourage the development of better algorithms. However, many methods developed with ATLAS v1.2 report low accuracy, are not publicly accessible or are improperly validated, limiting their utility to the field. Here we present ATLAS v2.0 (N = 1271), a larger dataset of T1w MRIs and manually segmented lesion masks that includes training (n = 655), test (hidden masks, n = 300), and generalizability (hidden MRIs and masks, n = 316) datasets. Algorithm development using this larger sample should lead to more robust solutions; the hidden datasets allow for unbiased performance evaluation via segmentation challenges. We anticipate that ATLAS v2.0 will lead to improved algorithms, facilitating large-scale stroke research.

Automatic Segmentation and Quantitative Assessment of Stroke Lesions on MR Images

“…Therefore, it was prone to learning any biases associated with the manual segmentation process [ 33 ]. Although the dataset was derived from multiple sites, the manual segmentation was done by a common group of tracers, which can introduce subjectivity and confounds [ 2 ].To overcome this, the model will need to be tested on independent datasets. This, however, represents another challenge as there are limited publicly available datasets for chronic stroke with manual segmentation labels [ 9 ].…”

“…Figure 1 is a visualization of the lesion overlap maps across all subjects ( n = 655) in the MNI space. 57.1% of subjects had at least one left hemisphere lesion, and 58.8% had at least one right hemisphere lesion [ 2 ].…”

“…In this quest, automatic lesion segmentation has emerged as a crucial tool [ 2 ]. Automatic image segmentation refers to assigning non-overlapping boundaries in an image to regions dissimilar in core features such as intensity or texture [ 3 ].…”

“…Automatic image segmentation refers to assigning non-overlapping boundaries in an image to regions dissimilar in core features such as intensity or texture [ 3 ]. There are numerous automatic segmentation algorithms for acute stroke; however, in comparison, there is a paucity of algorithms for chronic stroke lesions [ 2 ]. The reasons for this are multifold.…”

Functional MRI in Radiology—A Personal Review

“…There is a remaining high interest among researchers in the usage of these methods in basic neuroscience, and the community for OHBM is young and integrates many new possibilities and tools. Especially, the possibility to integrate datasets from different centers, for instance, by the ENIGMA interest groups [ 105 ], can overcome discrepant results from different samples due to a lack of power, particularly for patient studies [ 106 ] and for developing evaluation strategies in large datasets [ 107 ]. However, funding organizations have to overcome the tradition of predominantly providing funding for new investigations rather than for more advanced evaluation procedures of larger, already existing datasets.…”