FS-MSA: A Few-Shot, Self-Prompting 3-D Medical Image Segmentation Algorithm

Overview

Machine learning (ML) models are emerging as promising tools to aid non-invasive detection of medical anomalies (e.g. malignant tumors). However, privacy concerns and time-consuming labeling create medical image data shortages, making training conventional, data-hungry ML models infeasible. Researchers recently developed Segment Anything (SA), a foundational model displaying potential for addressing this problem. However, current implementations of SA cannot be reliably applied to most medical imaging, especially given the difficulty of small target recognition in tasks like semantic 3D tumor segmentation and the requirement of manual point or box input. SA’s enormous parameter space also complicates fine-tuning with low sample size. This study aims to address these problems by proposing a novel approach for few-shot medical image segmentation: "FS-MSA" (few-shot self-prompting anything). FS-MSA extracts boxes using a lightweight object detection architecture, using generated Regions of Interest to automatically prompt SA without manual input and improve small object localization. SA is also optimized using LoRA (Low-Rank Adaptation of Large Language Models) to reduce parameters. BraTS 2020, a publicly available MRI brain tumor dataset, is used as a proof-of-concept for this algorithm. This novel approach achieves state-of-the-art performance relative to existing approaches and shows promise for other few-shot medical segmentation tasks.

Competition history

• AJAS 2025 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2025 project