Abstract
Purpose : Clustering gaze patterns to assess visual search behavior is a complex process requiring a degree of subjectivity. This study tested an objective framework for clustering raw gaze data to extract the visual search strategies of clinicians screening images for disease.
Methods : Eye tracking data from a Gazepoint GP3 HD were collected from 26 expert participants (optometrists, general ophthalmologists and vitreoretinal specialists) as they graded diabetic retinopathy in up to 40 posterior pole fundus photographs, and their results grouped into subsets on the basis of correct and incorrect diagnosis. Hidden Markov Models were used to determine the areas of interest common to all images based on raw gaze data, which were manually annotated and expanded to account for hardware accuracy. Fixation data were then converted to directed acyclic graphs using these annotations and simplified with depth-first search to create fixation strings. Results were compared using Levenshtein distance (1) and clustered using Affinity Propagation (2), with exemplars generated to visualise each cluster.
Results : Hidden Markov Models clustered gaze data around the optic disc, macula, and superior and inferior arcades. Image annotations were created for the optic disc, superior arcade, inferior arcade, and nasal and temporal macular regions. Fixation strings for correct (n=4584) and incorrect (n=914) records were clustered and exemplars extracted for the largest clusters in each subset. The exemplar for the largest cluster in the correct subset (n=36) was [nasal macula, inferior arcade, superior arcade, temporal macula, nasal macula]. The largest cluster in the incorrect subset (n=26) was [temporal macula, nasal macula, inferior arcade, nasal macula, inferior arcade, temporal macula, superior arcade, temporal macula].
Conclusions : Our results show that an objective framework can be used to convert raw gaze data into groups of similar fixation strings. If confirmed by studies in other domains such as x-ray analysis, this framework may offer an automated and objective method for extracting gaze behavior in visual search tasks.
1. Levenshtein VI. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady 1966 Feb 10 (Vol. 10, No. 8, pp. 707-710).
2. Frey BJ, Dueck D. Clustering by passing messages between data points. science. 2007 Feb 16;315(5814):972-6.
Methods : Eye tracking data from a Gazepoint GP3 HD were collected from 26 expert participants (optometrists, general ophthalmologists and vitreoretinal specialists) as they graded diabetic retinopathy in up to 40 posterior pole fundus photographs, and their results grouped into subsets on the basis of correct and incorrect diagnosis. Hidden Markov Models were used to determine the areas of interest common to all images based on raw gaze data, which were manually annotated and expanded to account for hardware accuracy. Fixation data were then converted to directed acyclic graphs using these annotations and simplified with depth-first search to create fixation strings. Results were compared using Levenshtein distance (1) and clustered using Affinity Propagation (2), with exemplars generated to visualise each cluster.
Results : Hidden Markov Models clustered gaze data around the optic disc, macula, and superior and inferior arcades. Image annotations were created for the optic disc, superior arcade, inferior arcade, and nasal and temporal macular regions. Fixation strings for correct (n=4584) and incorrect (n=914) records were clustered and exemplars extracted for the largest clusters in each subset. The exemplar for the largest cluster in the correct subset (n=36) was [nasal macula, inferior arcade, superior arcade, temporal macula, nasal macula]. The largest cluster in the incorrect subset (n=26) was [temporal macula, nasal macula, inferior arcade, nasal macula, inferior arcade, temporal macula, superior arcade, temporal macula].
Conclusions : Our results show that an objective framework can be used to convert raw gaze data into groups of similar fixation strings. If confirmed by studies in other domains such as x-ray analysis, this framework may offer an automated and objective method for extracting gaze behavior in visual search tasks.
1. Levenshtein VI. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady 1966 Feb 10 (Vol. 10, No. 8, pp. 707-710).
2. Frey BJ, Dueck D. Clustering by passing messages between data points. science. 2007 Feb 16;315(5814):972-6.
| Original language | English |
|---|---|
| Article number | 5179 |
| Pages (from-to) | 1-1 |
| Number of pages | 1 |
| Journal | Investigative Ophthalmology & Visual Science |
| Volume | 65 |
| Issue number | 7 |
| Publication status | Published - 17 Jun 2024 |
| Externally published | Yes |
| Event | ARVO Annual Meeting 2024: Vision for the Future - Seattle, United States Duration: 5 May 2024 → 9 May 2024 |
