BoxCell: Leveraging SAM for Cell Segmentation with Box Supervision.

Area of Science:

• Digital pathology
• Medical image analysis
• Computational biology

Background:

• Accurate cell segmentation in histopathology images is crucial for disease diagnosis and treatment.
• Manual annotation is time-consuming and requires specialized medical expertise, hindering supervised learning.
• Weakly supervised learning, utilizing bounding box annotations, offers a viable alternative.

Purpose of the Study:

• To develop a novel cell segmentation framework, BoxCell, for histopathological images using weakly supervised learning.
• To leverage the pre-trained Segment Anything Model (SAM) without finetuning for cell segmentation.
• To improve the efficiency and accuracy of cell segmentation in medical imaging.

Main Methods:

• BoxCell utilizes SAM's ability to interpret bounding boxes as prompts for generating pseudo-masks during training.
• A standalone segmenter is trained using these pseudo-masks.
• At test time, BoxCell combines masks from the trained segmenter and SAM (prompted by an object detector) using integer programming.

Main Results:

• BoxCell significantly outperforms existing box-supervised image segmentation models on three public datasets (CoNSep, MoNuSeg, TNBC).
• The framework achieves substantial Dice score gains of 6-10 points.
• The proposed integer programming approach effectively reconciles complementary segmentation masks.

Conclusions:

• BoxCell offers a powerful and efficient solution for cell segmentation in histopathology using only bounding box supervision.
• The direct application of SAM without finetuning demonstrates its versatility in medical image analysis.
• This approach reduces the reliance on extensive manual annotation, making advanced segmentation more accessible.