Segment Anything Model (SAM) — Research Project

Research project focused on understanding SAM comprehensively and improving its performance for medical image segmentation.

Contents

File	Description
2304.02643v1.pdf	Original SAM paper (Kirillov et al., Meta AI, April 2023)
SAM_Comprehensive_Analysis.md	Architecture, training, limitations, and medical improvement strategies
SAM_Mask_Decoder_Deep_Dive.md	Detailed walkthrough of the mask decoder
MedSAM_Analysis.md	MedSAM findings, limitations, and research opportunities
SAM2_Colab_Notebook.ipynb	Interactive Colab notebook — load SAM2 from 🤗 Hugging Face

Quick Start

1. Upload SAM2_Colab_Notebook.ipynb to Google Colab
2. Runtime → Change runtime type → GPU (T4)
3. Run all cells

The notebook loads SAM2.1 models directly from Hugging Face (facebook/sam2.1-hiera-{tiny,small,base-plus,large}) — no manual downloads needed.

Research Focus

• Problem: SAM was trained on SA-1B (natural images only) → poor performance on medical images
• Approach: LoRA fine-tuning, adapters, multi-scale features, automated prompting, test-time augmentation
• Target modalities: CT, MRI, X-ray, pathology, ultrasound

Key References

• SAM Paper — Segment Anything (2023)
• SAM2 Paper — Segment Anything in Videos (2024)
• MedSAM — Segment Anything in Medical Images (Nature Communications, 2024)
• SAM2 on GitHub
• MedSAM on GitHub