Zero-Shot Evaluation and Domain-Specific Fine-Tuning of SAM 3 for Cataract Surgery Scenes
Abstract
We evaluate SAM 3 on the Cataract-1K dataset using four text-prompted configurations and a fixed 727-image test set: pretrained zero-shot (PTZS), domain-specific fine-tuning (DSFT), rare-class-augmented fine-tuning (RCAF), and contrast-refined fine-tuning (CRFT). PTZS achieves a micro presence F1 of 0.398 and a mean intersection over union (IoU) of 0.229, whereas all three domain-adapted configurations reach micro F1 values of approximately 0.95 and mean IoU values above 0.87. DSFT achieves a micro F1 of 0.955, RCAF provides the highest recall and mean IoU (0.885), and CRFT achieves the highest macro F1 (0.827) while producing the fewest false positives. These findings show that domain-specific adaptation is necessary for reliable text-conditioned segmentation of cataract-surgery scenes.