MedLSAM: Localize and Segment Anything Model for 3D Medical Images

Key Features

Foundation Model for 3D Medical Image Localization: MedLAM: MedLSAM introduces MedLAM as a foundational model for the localization of 3D medical images.
First Fully-Automatic Medical Adaptation of SAM: MedLSAM is the first complete medical adaptation of the Segment Anything Model (SAM). The primary goal of this work is to significantly reduce the annotation workload in medical image segmentation.
Segment Any Anatomy Target Without Additional Annotation: MedLSAM is designed to segment any anatomical target in 3D medical images without the need for further annotations, contributing to the automation and efficiency of the segmentation process.

Updates

2024.1.9: Release the training code
2023.10.15: Accelerate the inference speed. Add Sub-Patch Localization (SPL).
2023.07.01: Code released.

Details

The Segment Anything Model (SAM) has recently emerged as a groundbreaking model in the field of image segmentation. Nevertheless, both the original SAM and its medical adaptations necessitate slice-by-slice annotations, which directly increase the annotation workload with the size of the dataset. We propose MedLSAM to address this issue, ensuring a constant annotation workload irrespective of dataset size and thereby simplifying the annotation process. Our model introduces a few-shot localization framework capable of localizing any target anatomical part within the body. To achieve this, we develop a Localize Anything Model for 3D Medical Images (MedLAM), utilizing two self-supervision tasks: relative distance regression (RDR) and multi-scale similarity (MSS) across a comprehensive dataset of 14,012 CT scans. We then establish a methodology for accurate segmentation by integrating MedLAM with SAM. By annotating only six extreme points across three directions on a few templates, our model can autonomously identify the target anatomical region on all data scheduled for annotation. This allows our framework to generate a 2D bounding box for every slice of the image, which are then leveraged by SAM to carry out segmentations. We conducted experiments on two 3D datasets covering 38 organs and found that MedLSAM matches the performance of SAM and its medical adaptations while requiring only minimal extreme point annotations for the entire dataset. Furthermore, MedLAM has the potential to be seamlessly integrated with future 3D SAM models, paving the way for enhanced performance.

MedLSAM Image Fig.1 The overall segmentation pipeline of MedLSAM.

Feedback and Contact

Email: lyc745307452@gmail.com

Get Started

Main Requirements

torch>=1.11.0
tqdm
nibabel
scipy
SimpleITK
monai

Installation

Create a virtual environment conda create -n medlsam python=3.10 -y and activate it conda activate medlsam
Install Pytorch
git clone https://github.com/openmedlab/MedLSAM
Enter the MedSAM folder cd MedLSAM and run pip install -e .

Download Model

Download MedLAM checkpoint, SAM checkpoint, MedSAM checkpoint and place them at checkpoint/medlam.pth, checkpoint/sam_vit_b_01ec64.pth and checkpoint/medsam_vit_b.pth

Inference

GPU requirement

We recommend using a GPU with 12GB or more memory for inference.

Data preparation

StructSeg Task1 HaN OAR
WORD (Request for access is required to download this dataset.)

Note: You can also download other CT datasets and place them any place you want. MedLSAM will automaticly apply the preprocessing procedure during the inference time, so please do not normalize the original CT images!!!

After downloading the datasets, you should sort the data into "support" and "query" groups. This does not require moving the actual image files. Rather, you need to create separate lists of file paths for each group.

For each group ("support" and "query"), perform the following steps:

Create a .txt file listing the paths to the image files.
Create another .txt file listing the paths to the corresponding label files.

Ensure that the ordering of images and labels aligns in both lists. These lists will direct MedLSAM to the appropriate files during the inference process.
The file names are not important, as long as the ordering of images and labels aligns in both lists.

Example format for the .txt files:

image.txt

/path/to/your/dataset/image_1.nii.gz
...
/path/to/your/dataset/image_n.nii.gz

label.txt

/path/to/your/dataset/label_1.nii.gz
...
/path/to/your/dataset/label_n.nii.gz

Config preparation

MedLAM and MedLSAM load their configurations from a .txt file. The structure of the file is as follows:

[data]
support_image_ls      =  config/data/StructSeg_HaN/support_image.txt
support_label_ls      =  config/data/StructSeg_HaN/support_label.txt
query_image_ls        =  config/data/StructSeg_HaN/query_image.txt
query_label_ls        =  config/data/StructSeg_HaN/query_label.txt
gt_slice_threshold    = 10
bbox_mode             = SPL
slice_interval        = 2
fg_class              = [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22]
seg_save_path         = result/npz/StructSeg
seg_png_save_path     = result/png/StructSeg

[vit]
net_type                = vit_b

[weight]
medlam_load_path   = checkpoint/medlam.pth
vit_load_path  = checkpoint/medsam_20230423_vit_b_0.0.1.pth

Each of the parameters is explained as follows:

support_image_ls: The path to the list of support image files. It is recommended to use between 3 and 10 support images.
support_label_ls: The path to the list of support label files.
query_image_ls: The path to the list of query image files.
query_label_ls: The path to the list of query label files.
`

MedLSAM

Install / Use

README