StructSAM: Structure-Aware Prompt Adaptation for Lung Cancer Lesion Segmentation in CT

Overview

StructSAM introduces a novel framework that integrates anatomical priors and 3D contextual information to enhance segmentation of lung cancer lesions in CT imaging. It outperforms existing models including SAM, MedSAM, and other state-of-the-art methods in accuracy and generalization across datasets.

Background

Lung cancer is the leading cause of cancer-related deaths globally, necessitating precise lesion delineation from CT scans for effective clinical management. Automated segmentation is challenged by subtle lesion boundaries, anatomical variability, and the volumetric nature of CT data. Traditional deep learning models like U-Net have limitations in generalizing across diverse clinical settings. The Segment Anything Model (SAM) has shown promise in natural images but faces significant hurdles in medical imaging due to lack of anatomical awareness and volumetric modeling.

Data Highlights

Key Findings

• StructSAM incorporates anatomical priors (e.g., organ masks, vesselness) into prompt generation, improving segmentation at low-contrast lesion boundaries.
• A lightweight 3D-aware adapter aggregates inter-slice context, enhancing volumetric continuity and reducing slice-wise inconsistencies.
• Domain-aware parameter-efficient fine-tuning (PEFT) enables robust generalization across different datasets and institutions.
• StructSAM outperforms SAM, MedSAM, and other state-of-the-art medical segmentation models in both accuracy and cross-organ generalization.
• Existing SAM adaptations lack explicit structural priors and volumetric modeling, limiting clinical applicability in lung cancer lesion segmentation.

Clinical Implications

StructSAM’s integration of anatomical structure and volumetric context addresses critical limitations of current segmentation models, enabling more reliable lung lesion delineation in CT scans. This advancement supports improved precision in diagnosis, staging, and treatment planning for lung cancer patients. Its robust generalization across institutions facilitates broader clinical deployment of automated segmentation tools.

Conclusion

StructSAM represents a significant step forward in adapting foundation models for medical imaging by embedding structural awareness and volumetric continuity. This approach enhances the clinical viability of automated lung cancer lesion segmentation, potentially improving patient outcomes through more accurate imaging assessments.

References

1. Ronneberger et al. 2015 -- U-Net: Convolutional Networks for Biomedical Image Segmentation
2. Kirillov et al. 2023 -- Segment Anything Model (SAM)
3. Zhou et al. 2023 -- MedSAM: Medical Pretraining of SAM
4. Wang et al. 2024 -- SAM 2 and Video Processing Extensions