Structural Entities Extraction and Patient Indications Incorporation for Chest X-ray Report Generation

The automated generation of imaging reports proves invaluable in alleviating the workload of radiologists. A clinically applicable reports generation algorithm should demonstrate its effectiveness in producing reports that accurately describe radiology findings and attend to patient-specific indications. In this paper, we introduce a novel method, Structural Entities extraction and patient indications Incorporation (SEI) for chest X-ray report generation. Specifically, we employ a structural entities extraction (SEE) approach to eliminate presentation-style vocabulary in reports and improve the quality of factual entity sequences. This reduces the noise in the following cross-modal alignment module by aligning X-ray images with factual entity sequences in reports, thereby enhancing the precision of cross-modal alignment and further aiding the model in gradient-free retrieval of similar historical cases. Subsequently, we propose a cross-modal fusion network to integrate information from X-ray images, similar historical cases, and patient-specific indications. This process allows the text decoder to attend to discriminative features of X-ray images, assimilate historical diagnostic information from similar cases, and understand the examination intention of patients. This, in turn, assists in triggering the text decoder to produce high-quality reports. Experiments conducted on MIMIC-CXR validate the superiority of SEI over state-of-the-art approaches on both natural language generation and clinical efficacy metrics.

• 2024-09-09, Upload the Poster
• 2024-09-19, Update the repository to make it easy.
• 2024-09-19, Update the generated reports for the MIMIC-CXR test set.

• torch==2.1.2+cu118
• transformers==4.23.1
• torchvision==0.16.2+cu118
• radgraph==0.09
• Due to the specific environment of RadGraph, please refer to knowledge_encoder/factual serialization. py for the environment of the structural entities approach.

You can download checkpoints of SEI as follows:

• For MIMIC-CXR, you can download checkpoints from Baidu Netdisk (its code is MK13) and huggingface .

• For MIMIC-CXR, you can download medical images from PhysioNet.

• You can download medical reports from Google Drive. Note that you can apply with your license of PhysioNet, and its toy case is in knowledge_encoder/case.json

1. Config RadGraph environment based on knowledge_encoder/factual_serialization.py

   ===================environmental setting=================

   Basic Setup (One-time activity)

   a. Clone the DYGIE++ repository from here. This repository is managed by Wadden et al., authors of the paper Entity, Relation, and Event Extraction with Contextualized Span Representations.

   git clone https://github.com/dwadden/dygiepp.git

   b. Navigate to the root of the repo in your system and use the following commands to set the conda environment:

   conda create --name dygiepp python=3.7
   conda activate dygiepp
   cd dygiepp
   pip install -r requirements.txt
   conda develop . # Adds DyGIE to your PYTHONPATH

   c. Activate the conda environment:

   conda activate dygiepp

   Notably, for our RadGraph environment, you can refer to knowledge_encoder/radgraph_requirements.yml.

2. Configure radgraph_path and ann_path in knowledge_encoder/see.py. The ann_path corresponds to the local file path to the medical reports. The radgraph_path can be obtained directly from PhysioNet.

3. Run the knowledge_encoder/see.py to extract the factual entity sequence for each report.

4. Finally, the annotation.json becomes mimic_cxr_annotation_sen.json that is identical to new_ann_file_name variable in see.py

1. Run bash pretrain_mimic_cxr.sh to pretrain a model on the MIMIC-CXR data (Note that the mimic_cxr_ann_path is mimic_cxr_annotation_sen.json).

1. Config --load argument in pretrain_inference_mimic_cxr.sh. Note that the argument is the pre-trained model from the first stage.

2. Run bash pretrain_inference_mimic_cxr.sh to retrieve similar historical cases for each sample, forming mimic_cxr_annotation_sen_best_reports_keywords_20.json (i.e., the mimic_cxr_annotation_sen.json becomes this *.json file).

1. Extract and preprocess the indication section in the radiology report.

   a. Config ann_path and report_dir in knowledge_encoder/preprocessing_indication_section.py, and its value is mimic_cxr_annotation_sen_best_reports_keywords_20.json. Note that report_dir can be downloaded from PhysioNet.

   b. Run knowledge_encoder/preprocessing_indication_section.py, forming mimic_cxr_annotation_sen_best_reports_keywords_20_all_components_with_fs_v0227.json

2. Config --load argument in finetune_mimic_cxr.sh. Note that the argument is the pre-trained model from the first stage. Furthermore, mimic_cxr_ann_path is mimic_cxr_annotation_sen_best_reports_keywords_20_all_components_with_fs_v0227.json

3. Download these checkpoints. Notably, the chexbert.pth and radgraph are used to calculate CE metrics, and bert-base-uncased and scibert_scivocab_uncased are pre-trained models for cross-modal fusion network and text encoder. Then put these checkpoints in the same local dir (e.g., "/home/data/checkpoints"), and configure the --ckpt_zoo_dir /home/data/checkpoints argument in finetune_mimic_cxr.sh

4. Run bash finetune_mimic_cxr.sh to generate reports based on similar historical cases.

1. You must download the medical images, their corresponding reports (i.e., mimic_cxr_annotation_sen_best_reports_keywords_20_all_components_with_fs_v0227.json), and checkpoints (i.e., SEI-1-finetune-model-best.pth) in Section Datasets and Section Checkpoints, respectively.

2. Config --load and --mimic_cxr_ann_patharguments in test_mimic_cxr.sh

3. Run bash test_mimic_cxr.sh to generate reports based on similar historical cases.

4. Results on MIMIC-CXR are presented as follows:

5. Next, the code for this project will be streamlined.

If you use or extend our work, please cite our paper at MICCAI 2024.

• R2Gen Some codes are adapted based on R2Gen.
• R2GenCMN Some codes are adapted based on R2GenCMN.
• MGCA Some codes are adapted based on MGCA.

[1] Chen, Z., Song, Y., Chang, T.H., Wan, X., 2020. Generating radiology reports via memory-driven transformer, in: EMNLP, pp. 1439-1449.

[2] Chen, Z., Shen, Y., Song, Y., Wan, X., 2021. Cross-modal memory networks for radiology report generation, in: ACL, pp. 5904-5914.

[3] Wang, F., Zhou, Y., Wang, S., Vardhanabhuti, V., Yu, L., 2022. Multigranularity cross-modal alignment for generalized medical visual representation learning, in: NeurIPS, pp. 33536-33549.