The effect of SAM segmentation

[boomboomboooo]

Hello! I visualized the results of sam segmentation and found that it often didn't do a good job of distinguishing diseased areas. What parameters can I adjust to improve the segmentation?
I would appreciate it if you could guide me.

[FangHeng]

Thank you for your interest in our work and for your thoughtful question.

The default model of SAM, which was primarily trained on general-domain images, often does not generalize well to medical images—especially gigapiexl WSIs. This can lead to suboptimal segmentation of pathological regions. Prior research has also shown that SAM’s segmentation performance on pathology slides is limited. Some recent work, such as MedSAM, has shown better performance in medical image segmentation, so you might consider trying it if your goal is more accurate clinical segmentation.
For those who still wish to experiment with the original SAM, it does provide a range of hyperparameters that can be adjusted within the SamAutomaticMaskGenerator class. Below is an example of the configuration we have used:

segement_mask_generator = SamAutomaticMaskGenerator(
    model=sam,
    points_per_side=32,
    pred_iou_thresh=0.6,
    stability_score_thresh=0.6,
    crop_n_layers=1,
    crop_n_points_downscale_factor=2,
    min_mask_region_area=0,
)

Adjusting values such as points_per_side, pred_iou_thresh, and stability_score_thresh can help refine the quality of the automatically generated masks. However, even with tuning, our experiments showed that direct segmentation results from SAM on gigapixel WSIs often remain suboptimal. Consequently, we use these masks only as prior cues—for example, to guide a downstream model or provide weak supervision—rather than relying solely on raw SAM outputs for final segmentation.
If you need more accurate segmentation or plan to use these results in a downstream pipeline, you may need to explore other SAM variants or even fine-tune domain-specific models on your own dataset. These approaches typically account for the characteristics of histopathological data and can better capture fine-grained features of diseased regions in WSIs. This GitHub project may help.

I hope these suggestions help you refine your experiments, and please feel free to let us know if you have further questions.