Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Skill Acquisition of Lesion Segmentation in Novice Tracers
Poster Session A, Friday, September 12, 11:00 am - 12:30 pm, Field House
This poster is part of the Sandbox Series.
Isabella Huynh1, Erin Yi1, Jeremy Yeaton1, Danielle Fahey2; 1UC Irvine, 2University of Alabama
Introduction: Lesion segmentation is an essential prerequisite to lesion-symptom mapping (LSM), a method that investigates the relationship between the location of brain damage and cognitive function. Lesion segmentation involves hand-tracing the outline of a brain lesion onto a three-dimensional magnetic resonance image (MRI). We have set up an independent working group to allow individuals to develop lesion segmentation skills using freely available software, without reliance on expert neuroradiologists or specialized trainers. To our knowledge, lesion segmentation is not included in the standard curriculum of any graduate programs or medical schools, making it an uncommon yet valuable skill. Despite its absence, there is high demand for accurate lesion segmentation, specifically within studies that include hundreds of subjects. Our aim here is to establish and identify the best practices for lesion segmentation training. Methods: Over 22 weeks, participants were assigned 11 lesion tracings. We pre-selected the 11 lesions from the openly-available USC ATLAS dataset, focusing on lesions in the left hemisphere, within the frontal and parietal lobes. We began with an initial training period in which members received a lesion-tracing guide and four MRI scans with corresponding gold-standard segmentation to familiarize themselves with the software and understand lesion delineation. Our group met after completing each segmentation to receive feedback with slice-by-slice visual comparisons of their tracings against the gold standard and their peers, which they could also access independently. We measured participants’ accuracy by calculating the Dice coefficient (DC) to capture the degree of similarity between participants’ tracings and the gold-standard segmentation. Results: Participants (N = 13) included a range of social science undergraduates, graduate students, and one individual with prior MRI experience. They had an average DC with the gold-standard segmentation of .60 (range: 0.48-0.71) across the first five scans. By Scan 4, the average DC rose to 0.71, indicating growth and improvement in segmentation skill. Notably, one participant exemplified the greatest improvement, with an initial 0.43 in Scan 1 to 0.79 in Scan 4. Participants completed a brief survey after each scan. They reported consistently finding ventricles and sulcal structures difficult to trace, with axial and coronal views preferred and sagittal views most difficult to interpret. Despite persistent challenges with cortical structures, most participants showed increased efficiency and improved confidence by Week 5. Conclusion: At the end of the training period, we will conduct a reliability test of trainees by having them segment lesions on 3 new samples, which will be compared against gold-standard tracings. By evaluating participants' performance using voxel overlap metrics (Dice coefficient), we expect training and feedback to improve lesion tracing accuracy over time and reduce variability both against the gold standard and among tracers. It is worth noting that we intentionally selected scans to increase in difficulty, which may artificially hide progress despite participant improvement. Our findings will support the development of standardized and accessible training resources for lesion tracing and pave the way to enhanced diagnostic precision to create targeted treatment interventions in both research and clinical settings for neurological conditions.
Topic Areas: Methods, Disorders: Acquired