Gland Segmentation Using SAM With Cancer Grade as a Prompt

TL;DR

This paper enhances gland segmentation in cancer diagnosis by integrating cancer grade prompts into the Segment Anything Model, achieving state-of-the-art results through model modifications and training strategies.

Contribution

It introduces a novel prompt-based approach using cancer grade information within SAM for improved gland segmentation and cancer classification.

Findings

Achieved state-of-the-art segmentation performance.

Utilized cancer grade prompts to improve model accuracy.

Enhanced segmentation results with model modifications.

Abstract

Cancer grade is a critical clinical criterion that can be used to determine the degree of cancer malignancy. Revealing the condition of the glands, a precise gland segmentation can assist in a more effective cancer grade classification. In machine learning, binary classification information about glands (i.e., benign and malignant) can be utilized as a prompt for gland segmentation and cancer grade classification. By incorporating prior knowledge of the benign or malignant classification of the gland, the model can anticipate the likely appearance of the target, leading to better segmentation performance. We utilize Segment Anything Model to solve the segmentation task, by taking advantage of its prompt function and applying appropriate modifications to the model structure and training strategies. We improve the results from fine-tuned Segment Anything Model and produce SOTA results using this approach.