Medical Imaging Segmentation with SAMed

SAMed: A Universal Medical Image Segmentation Model using a Promptable Foundation Model (SAM)

Project Overview:
I am seeking an experienced AI/ML developer or deep learning engineer to help modify and implement the SAMed algorithm (Segment Anything Model adapted for medical imaging) for pancreas segmentation tasks across CT and MRI datasets.

We have already completed the baseline version of this project using nnUNet. The datasets are fully available and preprocessed, and all existing performance benchmarks have been recorded. Now, we want to replace the nnUNet-based approach with the SAMed algorithm to test its generalization, performance, and scalability in a medical context.


---

What’s Already Done:

All datasets (multi-center CT & MRI) are collected and organized (public and private).

The baseline segmentation system using nnUNet is implemented and evaluated.

Segmentation metrics (Dice, Jaccard, HD95, ASSD) have been calculated for the nnUNet results.

Documentation and initial pipeline are available.



---

Key Responsibilities:

Implement or fine-tune the SAMed algorithm, adapting SAM for medical image segmentation.

Replace nnUNet in our existing pipeline with SAMed while maintaining compatibility with preprocessing and evaluation scripts.

Handle DICOM and NIfTI medical image formats, ensuring proper handling of CT and MRI images.

Evaluate SAMed’s performance on the same datasets and compare it with the existing nnUNet results.

(Optional) Provide a simple GUI or interactive notebook for visualizing segmentation outputs.



---

Preferred Expertise:

Strong experience with PyTorch and transformer models.

Practical understanding of SAM (Segment Anything Model) and its architecture.

Familiarity with medical imaging workflows using tools like MONAI, SimpleITK, or NiBabel.

Hands-on experience in medical segmentation projects and model evaluation.

Bonus: Knowledge of foundation model fine-tuning or transfer learning.



---

Tools & Technologies:

Python, PyTorch

SAM / SAMed (Segment Anything Model for medical imaging)

MONAI, SimpleITK, NiBabel

Google Colab / Jupyter Notebook

GitHub for code sharing and version control



---

Deliverables:

Working SAMed-based segmentation pipeline

Replacement of nnUNet with SAMed in our current project

Evaluation report comparing SAMed vs nnUNet results

Pretrained model weights and source code

Documentation with instructions on training and testing

(Optional) A basic GUI or notebook demo interface