Introduction
Barrett’s dysplasia is often subtle, with poorly defined margins that are difficult to delineate endoscopically. Current computer-aided detection (CADe) systems demonstrate only moderate performance in accurately outlining lesion boundaries (segmentation). Improving segmentation accuracy is key to guiding precise biopsy and may potentially aid in better delineation of resection margins.
Aims & Methods
We aim to evaluate the performance of the Segment-Anything-Model 2 (SAM2), a state-of-the-art video segmentation model for delineating Barrett’s dysplastic lesions, when integrated with a CADe model.
Methods: We used a Barrett’s dataset comprising 250 patients, 913 images, and 29,219 video frames (111 videos). The test set included 9,550 video frames and 191 images, with dysplasia present in 5,570 frames and 79 images. Dysplastic areas were annotated by three experts with histological confirmation. A baseline CADe model (YOLO26m-seg) was trained for lesion detection and segmentation, and evaluated on both images and video frames. A CADe model is required to detect suspected dysplasia and provide an initial prompt to activate SAM2 to segment lesions in video frames.
We evaluated the following three approaches on the video frames from the testing dataset: (1) Baseline CADe model, (2) CADe-prompted SAM2 segmentation, and (3) Expert-prompted SAM2 segmentation. Model performance was evaluated using accuracy, sensitivity, and specificity. Segmentation performance was assessed using the mean intersection over union (mIoU), which measures overlap between predicted and true lesion areas, and the mean Dice coefficient, which reflects pixel-level similarity between two segmentations.
Results
The baseline CADe model achieved an accuracy of 0.75, sensitivity of 0.80, specificity of 0.68, mIoU of 0.59, and mDice of 0.63 when validated against images and videos. When validated against video frames only, the CADe model's performance was similar, with an accuracy of 0.75, sensitivity of 0.83, specificity of 0.65, mIoU of 0.58, and mDice of 0.63. CAD-prompted SAM2 segmentation achieved a better performance, with an accuracy of 0.83, sensitivity of 0.89, specificity of 0.64, mIoU of 0.65, and mDice of 0.73. Expert-prompted SAM2 segmentation had the best performance, with an accuracy of 0.97, sensitivity of 0.96, specificity of 0.99, mIoU of 0.77, and mDice score of 0.82.
Table 1: Performance of baseline CADe model against CADe and Expert-prompted SAM2 segmentation
Comparator | Accuracy | Sensitivity | Specificity | mIOU | mDice |
1) Baseline CADe model | 0.75 | 0.83 | 0.65 | 0.58 | 0.63 |
2) CADe-prompted SAM2 segmentation | 0.83 | 0.89 | 0.64 | 0.65 | 0.73 |
3) Expert-prompted SAM2 segmentation | 0.97 | 0.96 | 0.99 | 0.77 | 0.82 |
Conclusion
SAM2 segmentation improves the delineation of Barrett’s dysplastic lesions compared with standard CADe alone. This enhanced segmentation may assist endoscopists by improving the precision of biopsy targeting and supporting more complete endoscopic resection. Expert-prompted SAM2 achieved the highest agreement with full expert annotations, suggesting it can enable accurate video annotation with substantially reduced expert input.
Disclosure
The authours do not have any conflict of interest to declare