Toward uncertainty-aware manual delineation of brain tumours using eye-tracking and image-derived features.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
Background and Purpose: Accurate tumour delineation is key in radiotherapy workflows. Concurrently, diffuse tumour types are inherently difficult to delineate, rendering contour uncertainty information highly valuable for downstream treatment planning. In this study, we investigated whether uncertainty in manually delineated tumour contours can be inferred from clinician behaviour (eye and mouse movements) and image-derived indicators. Materials and Methods: In a two-stage controlled experiment, 36 clinical imaging experts described and manually delineated brain tumours on T2-FLAIR (fluid-attenuated inversion recovery) MRI scans, while their eye and mouse movements were recorded. Inter-observer contour variability was used as a proxy for contour uncertainty. In addition, we extracted image-derived features, including U-Net saliency maps, pixel-wise U-Net segmentation probabilities, and image entropy. To assess and predict contour uncertainty, we applied mixed-effects models and machine learning algorithms. Results: Higher contour uncertainty was associated with higher saccade velocities and greater fixation density. Uncertainty was also significantly correlated with segmentation error and image-derived features (p < 0.05). A random forest regressor, combining behavioural and image-derived features, explained 39% of the variance in contour uncertainty. Notably, features derived from downsampling layers were the strongest predictors, despite displaying the lowest saliency overlap with human attention. Conclusions: Our results suggest that uncertainty in manually delineated tumour contours can be estimated using behavioural and image-derived features, without explicit uncertainty annotations. While clinical deployment will require validation under realistic acquisition conditions and with specialty experts, these findings established the feasibility of passively inferred uncertainty as a foundation for future uncertainty-aware delineation tools.