Skip to content

[AAAI'26] PET2Rep: Towards Vision-Language Model-Drived Automated Radiology Report Generation for Positron Emission Tomography

Notifications You must be signed in to change notification settings

YichiZhang98/PET2Rep

Repository files navigation

PET2Rep

📖 arXiv | 🎈 GitHub | 🤗 HuggingFace

This repository is the official repo of PET2Rep: Towards Vision-Language Model-Drived Automated Radiology Report Generation for Positron Emission Tomography accepted by AAAI'26 .

image

Overview

PET2Rep is the first benchmark for evaluating VLMs in PET radiology report generation. It includes 565 whole-body PET/CT image-report pairs covering dozens of organs with metabolic information, which is essential for reflecting the real-world clinical comprehensiveness of PET imaging in oncologic diagnostics. The dataset is public available at here.

image

Quick Start

To get started, please clone this repository and follow the instructions below to set up your environment and run the demo.

git clone https://github.com/YichiZhang98/PET2Rep.git
cd PET2Rep

Create and activate a virtual environment.

conda create -n pet2rep python=3.10 -y
conda activate pet2rep

Here we provide an inference example for Qwen2.5-VL, other VLMs were also performed obeying the instruction of their technical reports. You can solve the environment by

pip install -r requirements.txt

Alternatively, you can also refer to the Hugging Face page of Qwen2.5-VL.

pip install git+https://github.com/huggingface/transformers accelerate
pip install qwen-vl-utils[decord]==0.0.8

Run the inference script of Qwen2.5-VL, inference_qwen25vl.py. And the content in the prompt file text_prompt/prompt_qwen25vl.jsonl will be fed into the model for inference.

python inference_qwen25vl.py

The results will be saved to the output file results/result_qwen25vl.jsonl.

Report Generation

We provide PET/CT slices as image input, and a typical nuclear medical report as an example for the VLM to genrate a structural report. The following is an demo case of the task:

Images:

slice1 slice2 slice3

Text (Chinese):

你是一位经验丰富的核医学科医师,擅长结合PET/CT图像和临床信息撰写诊断报告。
请你根据提供的全身FDG-PET/CT图像,按照结构化格式生成核医学报告。基于模版修改,需详细描述异常18F-FDG摄取病灶的位置、大小和程度(增高/降低/无异常)等。同时给出PET异常区域异常区域的CT所见。以下是参考的结构化报告输出模板,请根据患者PET/CT影像调整[]和【】中结论。\n参考结构化报告示例如下,严格按照以下格式输出:\n禁食状态下静脉注射18F-FDG,静息60分钟后行全身PET断层显像,全身断层影像显示:脑形态、结构[正常],双侧大脑皮质放射性分布【均匀、对称】,FDG【摄取未见明显异常】。各脑室、脑沟、脑裂、脑池[形态增宽,两侧脑室对称,中线结构无偏移]。\n双侧眼球形态和轮廓[无异常],FDG【无异常摄取】。右侧上颌窦黏膜[稍增厚伴部分钙化],FDG【摄取未见异常】。鼻咽壁[未见增厚,两侧腭扁桃体对称],FDG【呈生理性摄取】。喉咽[形态、结构未见异常]。双侧腮腺、颌下腺[形态、密度正常],FDG【生理性摄取】。甲状腺[形态、大小正常,密度均匀],FDG【摄取未见异常】。双侧颈深间隙、颌下[见数枚小淋巴结显示],部分淋巴结FDG【摄取增高】。\n两肺[透亮度增加,两肺上叶胸膜下见囊状透亮影。右肺中叶体积缩小,右肺中叶肺门旁见片状密度增高影],FDG【摄取增高】。心影[未见异常。部分动脉管壁钙化(包括冠状动脉)]。\n食管[未见扩张,管壁未见明显增厚或肿块],FDG【摄取未见增高】。肝缘[欠光整,肝裂未见明显增宽,CT平扫肝右叶下段见巨大稍低密度肿块影,边界欠清],FDG【摄取增高】。肝左叶[见稍低密度结节,边界尚清晰],FDG【摄取较本底摄取减低】。肝右叶[膈顶区另见结节状致密影]。门静脉主干及左右支[明显增粗,密度增高],FDG【摄取增高】。门腔间隙、肝胃间隙、腹膜后[见淋巴结显示],FDG【摄取轻度增高】。\n胆囊[稍饱满,胆囊壁未见增厚],局部FDG【摄取未见异常】。胰腺[形态正常,胰体部见小结节状钙化灶,主胰管未见增宽],FDG【摄取未见明显异常】。脾脏[形态、大小密度未见异常],FDG【摄取未见异常】。双侧泌尿系统【见对比剂残留影】;左肾[体积缩小,左肾见多枚囊状低密度影及稍高密度小结节,边界清],FDG【摄取缺失】;右肾[上极见一脂肪密度结节],FDG【摄取未见异常】。双侧肾盂肾盏及输尿管[未见增宽]。双侧肾上腺显影[未见明显异常],FDG【摄取未见异常】。\n胃[充盈尚可,胃壁未见明显增厚],部分胃壁FDG【摄取增高】。肠道[充盈欠佳,肠壁未见明显增厚或肿块],FDG【呈生理性摄取】。前列腺[稍增大,实质内见数枚钙化灶],FDG【摄取未见异常增高】。膀胱[充盈欠佳],内【见对比剂残留影】。\n骨质密度[不均匀减低],脊柱[部分椎体边缘骨质增生]。FDG【摄取未见异常】。

Text (English):

You are an experienced nuclear medicine physician, skilled in combining PET/CT images with clinical information to prepare diagnostic reports.
Please generate a nuclear medicine report in a structured format based on the provided whole-body FDG-PET/CT images. Based on the template modifications, please provide a detailed description of the location, size, and degree (increased/decreased/abnormal) of abnormal 18F-FDG uptake. Also provide the CT findings of the abnormal PET and CT areas. Below is a sample structured report output template. Please adjust the conclusions in [] and 【】 based on the patient's PET/CT images. The following is a sample structured report output, strictly following the following format: A whole-body PET scan is performed after 60 minutes of rest after intravenous administration of 18F-FDG in the fasting state. The whole-body PET scan shows: [normal] brain morphology and structure, 【uniform and symmetrical】 radioactivity distribution in the bilateral cerebral cortex, and 【no significant abnormalities】 in FDG uptake. The ventricles, sulci, fissures, and cisterns show [widening of the morphology, symmetry of the ventricles, and no deviation from the midline]. The morphology and contours of the bilateral eyes show [no abnormalities], and 【no abnormal】 FDG uptake. The right maxillary sinus mucosa shows [slight thickening with some calcification], and FDG uptake is 【normal】. The nasopharyngeal wall shows [no thickening, and the palatine tonsils are symmetrical]. FDG uptake is 【physiological】. The hypopharynx shows [no abnormalities in morphology or structure]. The bilateral parotid and submandibular glands show [normal morphology and density], and FDG uptake is 【physiological】. The thyroid gland shows [normal morphology and size, with uniform density], and FDG uptake is 【normal】. [Several small lymph nodes are visible] in the deep cervical spaces and submandibular spaces, with 【increased】 FDG uptake in some lymph nodes. Both lungs show [increased luminosity, with cystic lucencies observed beneath the pleura of both upper lobes. The right middle lobe is reduced in size, and a patchy increased density is observed near the hilum of the right middle lobe]. FDG uptake is 【increased】. The cardiac shadow shows [no abnormalities. Calcification of some arterial walls (including the coronary arteries) is observed]. The esophagus shows [no dilatation, no significant wall thickening or mass], and 【no increased】 FDG uptake. The liver margins are [not smooth, and the fissures are not significantly widened. A large, slightly hypodense mass with unclear margins is observed in the lower segment of the right liver lobe on plain scan CT]. 【Increased】 FDG uptake is observed. [A slightly hypodense nodule with a well-defined margin] is observed in the left liver lobe. FDG uptake is 【reduced compared to background uptake】. [A nodular, dense shadow is also observed in the diaphragmatic dome region] of the right liver lobe. The main portal vein and its left and right tributaries show [significant thickening and increased density], with 【increased】 FDG uptake. [Lymph nodes are observed] in the portal space, hepatogastric space, and retroperitoneum, with 【mildly increased】 FDG uptake. The gallbladder is [slightly full, with no thickening of the gallbladder wall]. Local FDG uptake is 【normal】. The pancreas shows [normal morphology, with small nodular calcifications in the pancreatic body and no widening of the main pancreatic duct]. FDG uptake shows 【no significant abnormalities】. The spleen shows [no abnormalities in morphology, size, or density]. FDG uptake shows 【no abnormalities】. The bilateral urinary system shows residual contrast agent. The left kidney shows [reduced size, with multiple cystic, low-density shadows and slightly high-density nodules with clear margins]. FDG uptake is 【absent】. The right kidney shows [a fatty nodule at the upper pole]. FDG uptake shows 【no abnormalities】. The renal pelvis, calyces, and ureters show [no widening]. The adrenal glands show [no significant abnormalities]. FDG uptake shows 【no abnormalities】. The stomach shows [adequate filling, with no significant thickening of the stomach wall]. 【Increased】 FDG uptake is observed in portions of the stomach wall. The intestine shows [poor filling, with no significant thickening or masses in the bowel wall]. FDG uptake is 【physiological】. The prostate is [slightly enlarged, with several calcifications within the parenchyma]. FDG uptake is 【normal】. The bladder is [poorly filled, with residual contrast agent visible]. Bone density is [unevenly decreased], and the spine has [bone hyperplasia at the margins of some vertebrae]. FDG uptake is 【normal】.

Citation

If you find this repository helpful, please consider citing:

@article{zhang2025pet2rep,
  title={PET2Rep: Towards Vision-Language Model-Drived Automated Radiology Report Generation for Positron Emission Tomography},
  author={Zhang, Yichi and Zhang, Wenbo and Ling, Zehui and Feng, Gang and Peng, Sisi and Chen, Deshu and Liu, Yuchen and Zhang, Hongwei and Wang, Shuqi and Li, Lanlan and others},
  journal={arXiv preprint arXiv:2508.04062},
  year={2025}
}

About

[AAAI'26] PET2Rep: Towards Vision-Language Model-Drived Automated Radiology Report Generation for Positron Emission Tomography

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages