We developed a domain-speciffc large language-vision assistant (PA-LLaVA) for pathology image understanding. Specifically, (1) we first construct a human pathology image-text dataset by cleaning the public medical image-text data for domainspecific alignment; (2) Using the proposed image-text data, we first train a pathology language-image pretraining (PLIP) model as the specialized visual encoder for pathology image, and then we developed scale-invariant connector to avoid the information loss caused by image scaling; (3) We adopt two-stage learning to train PA-LLaVA, first stage for domain alignment, and second stage for end to end visual question & answering (VQA) task.
[24/08/30] The test code has been update!
[24/08/20] Model weights have been updated in HuggingFace! OpenFace-CQUPT/Pathology-LLaVA
The weights for PLIP and the weights for both the domain alignment and instruction fine-tuning phases of PA-LLaVA are disclosed in the HuggingFace(https://huggingface.co/OpenFace-CQUPT/Pathology-LLaVA).
These public datasets contain substantial amounts of data unrelated to human pathology. To obtain the human pathology image-text data, we performed two cleaning processes on the raw data, as illustrated in the follow figture: (1) Removing nonpathological images. (2) Removing nonhuman pathology data. Additionally, we excluded image-text pairs with textual descriptions of fewer than 20 words. Ultimately, we obtained 518,413 image-text pairs (named "PCaption-0.5M" ) for the aligned training dataset.
Instruction fine-tuning phase we only cleaned PMC-VQA in the same way and obtained 15,788 question-answer pairs related to human pathology. Lastly, we combined PathVQA and Human pathology data obtained from PMC-VQA, thereby constructing a dataset of 35543 question-answer pairs data.
Here we only provide the download link for the public dataset and expose the image id index of our cleaned dataset on HuggingFace(https://huggingface.co/OpenFace-CQUPT/Pathology-LLaVA).
PubMedVision-Alignment: FreedomIntelligence/PubMedVision · Datasets at Hugging Face
PMC-OA: axiong/pmc_oa · Datasets at Hugging Face
Quilt-1M: Quilt-1M: One Million Image-Text Pairs for Histopathology (zenodo.org)
PathVQA: https://drive.google.com/drive/folders/1G2C2_FUCyYQKCkSeCRRiTTsLDvOAjFj5
PMC-VQA: xmcmic/PMC-VQA · Datasets at Hugging Face
ICIAR 2018 BACH: https://iciar2018-challenge.grand-challenge.org/Download/
OSCC: https://data.mendeley.com/datasets/ftmp4cvtmb/1
ColonPath : https://medfm2023.grand-challenge.org/datasets
First, use the image index of the clean dataset provided by us to extract the human pathological dataset, and then process it into the following format:
[
{
"image": ,
"caption":
},
]
Finally, run dataformate.py to get the format needed to train the model.
python dataformat.py
We used xtuner as a training tool, so please go to xtuner official to complete the environment configuration [https://github.com/InternLM/xtuner]. Then add the xtuner_add/pallava file to the installed xtuner code with the following location structure.
NPROC_PER_NODE=8 NNODES=2 PORT=12345 ADDR= NODE_RANK=0 xtuner train pallava_domain_alignment.py --deepspeed deepspeed_zero2 --seed 1024
NPROC_PER_NODE=8 NNODES=2 PORT=12345 ADDR= NODE_RANK=0 xtuner train pallava_instruction_tuning.py --deepspeed deepspeed_zero2 --seed 1024
First, replace or add all the files in xtuner_add/tool_add into the tool file of the xtuner runtime file with the following file location structure:
Our released weights are distributed training weights that can be directly loaded for training through XTuner. If you need merged weights, they can be merged using XTuner (using the weights from the domain alignment phase as an example):
If you need to test the caption inference task with the first stage weights：
xtuner convert pth_to_hf path/pallava_domain_alignment.py ./domain_alignment_weight.pth ./domain_alignment_weight_ft
xtuner convert merge meta-llama/Meta-Llama-3-8B-Instruct ./domain_alignment_weight_ft/llm_adapter ./domain_alignment_weight_ft/llm_merge_lora
If you need to use phase 2 weights for classification or VQA tasks(The VQA question-and-answer style has been provided in the paper):
xtuner convert pth_to_hf path/pallava_instruction_tuning.py ./instruction_tuning_weight.pth ./instruction_tuning_weight_ft
xtuner convert merge meta-llama/Meta-Llama-3-8B-Instruct ./instruction_tuning_weight_ft/llm_adapter ./instruction_tuning_weight_ft/llm_merge_lora
NPROC_PER_NODE=8 xtuner pathvqa meta-llama/Meta-Llama-3-8B-Instruct --visual-encoder PLIP --llava ./instruction_tuning_weight_ft --prompt-template llama3_chat --data-path absolute_path/Path_VQA/path_vqa_test.json --work-dir absolute_path/logs/pathvqa --launcher pytorch --anyres-image
NPROC_PER_NODE=8 xtuner pmcvqa meta-llama/Meta-Llama-3-8B-Instruct --visual-encoder PLIP --llava ./instruction_tuning_weight_ft --prompt-template llama3_chat --data-path absolute_path/PMC-VQA/pmc-vqa_test_clean_answer_abcd.json --work-dir absolute_path/logs/pmcvqa --launcher pytorch --anyres-image
Here is an example with OSCC data.
NPROC_PER_NODE=8 xtuner zero_shot meta-llama/Meta-Llama-3-8B-Instruct --visual-encoder PLIP --llava ./instruction_tuning_weight_ft --prompt-template llama3_chat --data-path absolute_path/OSCC/oscc.json --work-dir absolute_path/logs/oscc --launcher pytorch --anyres-image
python test/f1.py
python python test/gpt4-scores.py
@INPROCEEDINGS{10821785,
author={Dai, Dawei and Zhang, Yuanhui and Xu, Long and Yang, Qianlan and Shen, Xiaojing and Xia, Shuyin and Wang, Guoyin},
booktitle={2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
title={PA-LLaVA: A Large Language-Vision Assistant for Human Pathology Image Understanding},
year={2024},
volume={},
number={},
pages={3138-3143},
keywords={Connectors;Pathology;Visualization;Codes;Computational modeling;Biological system modeling;Data models;Cleaning;Bioinformatics;Biomedical imaging;Pathology Image Understanding;VQA;LLaVA},
doi={10.1109/BIBM62325.2024.10821785}}
This repo is currently maintained by Dawei Dai (dw_dai@163.com) and his master's student Yuanhui Zhang (S230233056@stu.cqupt.edu.cn).