Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
This is an official repo for fine-tuning SAM to customized medical images.
| Date | Stars |
|---|---|
| 2026-07-31 | 280 |
| 2026-08-03 | 280 |
| 2026-08-06 | 280 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Finetune SAM on your customized medical imaging dataset Authors: [Hanxue Gu*](https://scholar.google.com/citations?hl=en&user=aGjCpQUAAAAJ&view_op=list_works&sortby=pubdate), [Haoyu Dong*](https://scholar.google.com/citations?user=eZVEUCIAAAAJ&hl=en), [Jichen Yang](https://scholar.google.com/citations?user=jGv3bRUAAAAJ&hl=en), [Maciej A. Mazurowski](https://scholar.google.com/citations?user=HlxjJPQAAAAJ&hl=en) **Notice:** 🥰Hi guys, since my github is not linked to my work email thus i might not reply to issues or questions quickly. Feel free to email me if you meet issues when using this repo, and i am glad to help. Here is my email: [email protected]. This is the official code for our paper: [How to build the best medical image segmentation algorithm using foundation models: a comprehensive empirical study with Segment Anything Model](https://arxiv.org/abs/2404.09957), where we explore three popular scenarios when fine-tuning foundation models to customized datasets in the medical imaging field: (1) only a single labeled dataset; (2) multiple labeled datasets for different tasks; and (3) multiple labeled and unlabeled datasets; and we design three common experimental setups, as shown in figure 1.  Our work summarizes and evaluates existing fine-tuning strategies with various backbone architectures, model components, and fine-tuning algorithms across 18 combinations, and 17 datasets covering all common radiology modalities.  Based on our extensive experiments, we found that: 1. fine-tuning SAM leads to slightly better performance than previous segmentation methods. 2. fine-tuning strategies that use parameter-efficient learning in both the encoder and decoder are superior to other strategies. 3. network architecture has a small impact on the final performance, 4. further training SAM with self-supervised learning can improve final model performance. To use our codebase, we provide (a) codes to fine-tune your medical imaging dataset on either automatic/prompt-based setting, (b) pretrained weights we got from Setup 3 using task-agnostic self-supervised learning, which we found as good pretrained weights instead of initial SAM providing better performance for downstream tasks. ## Bug fixes: - [X] May-10-2024, fixed the bug that when we updated the dataset.py at May 6th for multi class support, the mask resize processing was accidently forgotten. - [X] May-10-2024, fixed the bug that the provided demo for single gpu trianing only support updating decoder but the image encoder's gradients were not calculated. - [X] June-10-2024, fixed the bug that cfg.py was not updated as the same version of train.sh which didn't include two configs as 'train_img_list' and 'val_img_list'. ## Updated functions: - [X] May-15-2024, add functions to auto save training args and load args for validation; save your time for manual definition. - [X] May-15-2024, add two jupyter-notebooks showing examples about how to make predictions on 3D volumes/2D pngs without ground truth; and for visualization. - [X] May-15-2024, provide two additional example demos. - [X] June-10-2024, add spatial transformation choice in dataset.py ## a): fine-tune to one single task-specific dataset ### Step 0: setup environment If using conda enviroment: ```bash conda env create -f environment.yml ``` If directly using pip ```bash pip install -r requirements.txt ``` ### Step 1: dataset preparation. Please prepare your images and mask pairs in 2D slices first. If your original dataset is in 3D format, ple
Excerpt of 12,932 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:f3b71cdc19eec6f8, desc:fine-tuning, desc:fine tuning, name:finetune
matched fp:f3b71cdc19eec6f8, topic:foundation-models