AbdomenAtlas/xx_update.py

#coding:utf-8
'''

write by ygq

create on 2025-8-18

update AbdomenAtlas3.0 data clean



https://arxiv.org/pdf/2407.16697

https://zhuanlan.zhihu.com/p/19339643417



AbdomenAtlas 3.0 是目前公开的最大规模腹部 CT 图像-文本配对数据集，旨在解决医学影像中的肿瘤检测与报告生成难题。

该数据集包含 9,262 例 3D CT 扫描，来源于 88 家医疗机构，覆盖 19 个国家，并且是首个提供逐像素（per-voxel）标注、详细肿瘤报告以及肿瘤分期信息的公开数据集。

这些 CT 扫描数据通过标准医学影像格式（NIfTI 和 DICOM）存储，具备体素间距及 HU 值等临床信息。AbdomenAtlas 3.0 整合并重新标注了 17 个公共数据集，经过 12 位放射科医生的审核，共标注了 8,562 个肿瘤实例，其中包括 3,036 个肝脏肿瘤、354 个胰腺肿瘤和 4,239 个肾脏肿瘤。此外，数据集包含 2,947 份肿瘤报告，其中 948 份为早期肿瘤报告（≤2 cm），260 份报告提供了胰腺肿瘤的 T 分期（T1-T4），并首次公开肝脏 8 个亚段和胰腺 3 个亚段的逐像素标注，以及肿瘤与关键血管（如 SMA、CA 等）的接触标注。

通过 RadGPT 自动生成的结构化和叙述性报告，数据集详细描述了肿瘤大小、形状、位置、体积以及与周围血管和器官的相互作用。这些报告的生成准确性经过验证，在检测小肿瘤（≤2 cm）方面，RadGPT 的敏感性/特异性显著优于现有方法（例如肝脏：80%/73%，胰腺：77%/77%）。数据集还包含 240 份“人类-AI 融合报告”，结合了放射科医生的临床笔记和 AI 的精确量化结果。AbdomenAtlas 3.0 的意义在于，它首次提供了一个全面的腹部 CT 图像-文本配对数据集，填补了公开领域中腹部肿瘤检测数据的空白，并为推动医学影像中的自动化肿瘤检测、分期和报告生成奠定了基础。这一数据集不仅在规模和多样性上领先，还通过结合 AI 和放射科医生的专业知识，提供了高质量的标注和诊断支持，将有助于提升 AI 模型在医学影像分析中的实际临床应用能力。



数据集统计信息

总数据量：

9,262 例 3D CT 扫描，来源于 88 家医疗机构，覆盖 19 个国家。

包含 8,562 个肿瘤实例：

肝脏肿瘤：3,036 个实例（929 份报告）

胰腺肿瘤：354 个实例（344 份报告）

肾脏肿瘤：4,239 个实例（1,674 份报告）

6,061 份无肿瘤报告（作为对照组）

小肿瘤（≤2 cm）：

943 份小肿瘤相关报告：

肝脏：347 个实例（占肝脏肿瘤的 37.4%）

胰腺：83 个实例（占胰腺肿瘤的 24.1%）

肾脏：466 个实例（占肾脏肿瘤的 27.8%）

肿瘤分期与解剖结构：

260 份胰腺肿瘤分期报告（T1–T4）

提供肝脏 8 个亚段和胰腺 3 个亚段（头、体、尾）的逐像素分割

标注了肿瘤与关键血管（如 SMA、CA、CHA 等）的接触角度

图像与文本配对：

1.8M 文本 Token，包含三类报告：

结构化报告：基于模板生成，提供定量信息（如肿瘤体积、位置等）

叙述性报告：通过 LLM 转换，模仿目标医院的报告风格

人类-AI 融合报告：240 份，结合临床笔记与 AI 生成的内容



AbomentAtlas数据集中每个病例里面的segmentions都是包含了25个器官组织的标注文件，同时也包含一个combined_labels.nii.gz的文件【里面加上背景值包含了0-25的数值

1	aorta

2	gall_bladder

3	kidney_left

4	kidney_right

5	liver

6	pancreas

7	postcava

8	spleen

9	stomach

10	adrenal_gland_left

11	adrenal_gland_right

12	bladder

13	celiac_trunk

14	colon

15	duodenum

16	esophagus

17	femur_left

18	femur_right

19	hepatic_vessel

20	intestine

21	lung_left

22	lung_right

23	portal_vein_and_splenic_vein

24	prostate

25	rectum





参考TotalSegment分别存储25个器官的label处理后的数据文件

'''
import os
import glob
import pandas as pd
import SimpleITK as sitk
import argparse
import json
from tqdm import tqdm
from util import meta_data
import util
import numpy as np
# from bert_helper import *

# model_name = "bert-large-uncased"
# reduce_method = 'mean'
# max_words_num = 32  # max number of words in the caption > 2
    
# embeder, tokenizer = get_frozen_embeder(model_name)

# string1 = "modality: ct, gender: female, age: 51, roi: abdomen"
# embeder_output1 = str2emb(string1, max_words_num, embeder, tokenizer, reduce_method=reduce_method)

# string2 = "modality: ct, gender: female, age: 50, roi: head"

# embeder_output2 = str2emb(string2, max_words_num, embeder, tokenizer, reduce_method=reduce_method)

# input_size = embeder.config.vocab_size
# in_size = embeder.config.hidden_size
    
# print(embeder, input_size, in_size)
# print(tokenizer)
    

# print(embeder_output1)
# print(embeder_output1.shape)  # torch.Size([1, 8, 768])
    
    
# print(embeder_output2)
# print(embeder_output2.shape)  # torch.Size([1, 8, 768])
    

# error = torch.abs(embeder_output1 - embeder_output2)
# print(error)
# print("Embedding distance between the two sentences: ")
# print(f"String1: {string1}")
# print(f"String2: {string2}")
# print(torch.mean(error))


# exit()


# meta_id_name='Patient'
# meta_weeks_name='Weeks'
# meta_fvc_name='FVC'
# meta_percent_name='Percent'
# meta_age_name='Age'
# meta_sex_name='Sex'
# meta_status_name='SmokingStatus'

TASK_VALUE="segmentation"
CLAMP_RANGE_CT = [-300,300]
CLAMP_RANGE_MRI = [-1,0] # MRI images threshold placeholder TBC...

##判定是否有效胸部的肺部体积阈值ml
LUNG_VOL_THRESH=1000
FEMUR_VOL_THRESH=80
KIDNEY_VOL_THRESH=100
gall_bladder_VOL_THRESH=12
ROI="abdomen"

PROCESS_FLAG=True

LABEL_DICT={
    "0":"backgroud",
    "1":"aorta",
    "2":"gall_bladder",
    "3":"kidney_left",
    "4":"kidney_right",
    "5":"liver",
    "6":"pancreas",
    "7":"postcava",
    "8":"spleen",
    "9":"stomach",
    "10":"adrenal_gland_left",
    "11":"adrenal_gland_right",
    "12":"bladder",
    "13":"celiac_trunk",
    "14":"colon",
    "15":"duodenum",
    "16":"esophagus",
    "17":"femur_left",
    "18":"femur_right",
    "19":"hepatic_vessel",
    "20":"intestine",
    "21":"lung_left",
    "22":"lung_right",
    "23":"portal_vein_and_splenic_vein",
    "24":"prostate",
    "25":"rectum"
}
# def find_metadata_files(path):
#     # for Cancer Image Archive (TCIA) dataset
#     search_pattern = os.path.join(path, '**', 'metadata.csv')
#     return glob.glob(search_pattern, recursive=True)

def find_metadata_files(path):
    # for Cancer Image Archive (TCIA) dataset
    search_pattern = os.path.join(path, '*.csv')
    return glob.glob(search_pattern, recursive=True)
##added by yanguoqing on 20250527
def find_image_dirs(path):
    return os.listdir(path)

##modify by yanguoqing on 20250527
def load_dicom_images(folder_path):
    reader = sitk.ImageSeriesReader()
    dicom_names = reader.GetGDCMSeriesFileNames(folder_path)
    reader.SetFileNames(dicom_names)
    image = reader.Execute()
    return dicom_names,image

##added by yanguoqing on 20250527
def load_dicom_tag(imgs):
    reader = sitk.ImageFileReader()
    # dicom_names = reader.GetGDCMSeriesFileNames(folder_path)
    reader.SetFileName(imgs)
    reader.ReadImageInformation()  # 仅读取元信息，不加载像素数据
    # metadata_keys = reader.GetMetaDataKeys()
    tag=reader.Execute()
    return tag

def load_nrrd(fp):
    return sitk.ReadImage(fp)

def save_nifti(image, output_path, folder_path):
    # Set metadata in the NIfTI file's header
    output_dirpath = os.path.dirname(output_path)
    if not os.path.exists(output_dirpath):
        print(f"Creating directory {output_dirpath}")
        os.makedirs(output_dirpath)
    # Set metadata in the NIfTI file's header
    image.SetMetaData("FolderPath", folder_path)
    sitk.WriteImage(image, output_path)

##modify by yanguoqing on 20250527
def convert_windows_to_linux_path(windows_path):
    # Replace backslashes with forward slashes and remove the drive letter
    # Some meta files have windows paths, but the data is stored on a linux server
    linux_path = windows_path.replace('\\', '/')
    if ':' in linux_path:
        linux_path = linux_path.split(':', 1)[1]
    return linux_path

def simpleitk_volume_calculation(image_path):
    """

    使用SimpleITK简化体积计算流程，计算肺部体积，左肺或右肺超过400即认定为有效throax

    """

    image=util.load_nifti(image_path)
    # 获取体素尺寸
    spacing = image.GetSpacing()
    voxel_volume = spacing[0] * spacing[1] * spacing[2]  # mm³
    
    # print(f"图像尺寸: {image.GetSize()}")
    # print(f"体素间距: {spacing}")
    # print(f"单个体素体积: {voxel_volume:.6f} mm³")
    ##计算有效像元数量
    image_array2 = sitk.GetArrayFromImage(image)
    valid_pxiels=image_array2[image_array2==1].sum()
    if valid_pxiels<10:
        return 0
    # 简单的阈值分割（需要根据实际情况调整阈值）
    segmented = sitk.BinaryThreshold(image, lowerThreshold=1, upperThreshold=1)
    
    # 统计体素数量
    statistics = sitk.LabelShapeStatisticsImageFilter()
    statistics.Execute(segmented)
    
    voxel_count = statistics.GetNumberOfPixels(1)
    volume_mm3 = voxel_count * voxel_volume
    volume_ml = volume_mm3 / 1000.0
    
    # print(f"体素数量: {voxel_count}")
    # print(f"器官体积: {volume_ml:.2f} mL")
    
    return volume_ml

def main(target_path, output_dir):
    metadata_files = find_metadata_files(target_path)
    pid_dirs=find_image_dirs(target_path)
    failed_files = []
    label_dict={}
    if not os.path.isdir(output_dir):
        os.makedirs(output_dir)
    json_output_path = os.path.join(output_dir, 'xx.json')
    failed_files_path = os.path.join(output_dir, 'yy.json')
    #meta = meta_data()
    with open(json_output_path,'r') as fi:
        fj=json.load(fi)
    '''

    # Initialize the JSON file

    if not os.path.exists(json_output_path):

        with open(json_output_path, 'w') as json_file:

            json.dump({}, json_file)

    '''
    if pid_dirs:
        for pid_dir in tqdm(pid_dirs, desc="Processing pid dirs"):
            if not os.path.isdir(os.path.join(target_path,pid_dir)):
                continue
            if not pid_dir.startswith("BDMAP_"):
                continue
            
            meta_file=os.path.join(target_path,'%s.csv'%pid_dir)
            if os.path.isfile(meta_file):
                mf_flag=True
            #    df_meta=pd.read_csv(meta_file,sep=',')
            else:
                mf_flag=False
            
            full_path=os.path.join(target_path,pid_dir,"ct.nii.gz")
            
        
            try:
                '''

                dicom_image=util.load_nifti(full_path)

                spacing_info = dicom_image.GetSpacing()

                print('SPACING INFO:', spacing_info)

                

                # metadata_keys = dicom_image.GetMetaDataKeys()

                

                # dtag=load_dicom_tag(dicom_fp[0])

                # uid=dtag.GetMetaData('0020|000e') ##Series Instance UID

                # modality=dtag.GetMetaData('0008|0060')##Modality

                uid=pid_dir

                modality="CT"

                study='AbdomenAtlas'##Dataset_name

                CIA_other_info = {

                'Study_UID':uid,

                'metadata_file':''

                # 'Series_Description':serise_desc

                }

                CIA_other_info['split'] = "train"

                if mf_flag:

                    CIA_other_info['metadata_file']=meta_file



                size = list(dicom_image.GetSize())

                resampler =util.get_unisize_resampler(dicom_image, interpolator='linear', spacing=spacing_info, size=size)



                # resize the image 

                if resampler is not None:

                    proces_image = resampler.Execute(dicom_image)

                    print('SPACIE INFO AFTER', proces_image.GetSpacing())

                    CIA_other_info['Resample'] = True

                else:

                    proces_image = dicom_image

                    CIA_other_info['Resample'] = False

                

                ##

                # CIA_other_info['Image_id']=meta_image_id

                # CIA_other_info['Weeks']=str(meta_weeks)

                # CIA_other_info['FVC']=str(meta_fvc)

                # CIA_other_info['Percent']=str(meta_percent)

                # CIA_other_info['Age']=str(meta_age)

                # CIA_other_info['Sex']=meta_sex

                # CIA_other_info['Smoke_Status']=meta_status

                # threshold the image

                if 'CT' in modality:

                    proces_image = util.clamp_image(proces_image, CLAMP_RANGE_CT)

                else:

                    pass



                output_path = os.path.join(output_dir,uid, f"{uid}.nii.gz")

                # output_path=convert_windows_to_linux_path(output_path)

                save_nifti(proces_image, output_path, full_path)

                print(f"Saved NIfTI file to {output_path}")

                '''
                ##segment
                label_path_dict = {}
                label_flag=True
                
                label_paths = os.path.join(target_path,pid_dir, 'segmentations')
                label_files=glob.glob("%s/*.nii.gz"%(label_paths))
                #print(label_paths,label_files)
                pelvis_flag=False
                thorax_flag=False
                lung_min=0
                lung_max=0
                kidney_flag=False
                gall_bladder_flag=False
                if len(label_files)>0:
                    for lf in label_files:
                        lf_name=os.path.basename(lf)
                        
                        lf_tissue=lf_name.replace(".nii.gz","")
                        
                        if 'femur' in lf_tissue:
                            vol_femur=simpleitk_volume_calculation(lf)
                            print(lf_tissue,vol_femur)
                            if vol_femur>=FEMUR_VOL_THRESH:
                                pelvis_flag=True
                        if 'lung' in lf_tissue:
                            vol_lung=simpleitk_volume_calculation(lf)
                            print(lf_tissue,vol_lung)
                            lung_max=max(lung_max,vol_lung)
                            if lung_min==0:
                                lung_min=vol_lung
                            else:
                                lung_min=min(lung_min,vol_lung)
                            if lung_min>=LUNG_VOL_THRESH:
                                thorax_flag=True
                        if 'kidney_right' in lf_tissue:
                            vol_kidney=simpleitk_volume_calculation(lf)
                            print(lf_tissue,vol_kidney)
                            if vol_kidney>=KIDNEY_VOL_THRESH:
                                kidney_flag=True

                        if 'gall_bladder' in lf_tissue:
                            vol_gall_bladder=simpleitk_volume_calculation(lf)
                            print(lf_tissue,vol_gall_bladder)
                            if vol_gall_bladder>=gall_bladder_VOL_THRESH:
                                gall_bladder_flag=True     
                        '''

                        label_image=load_nrrd(lf)

                        resampler =util.get_unisize_resampler(label_image, interpolator='nearest', spacing=spacing_info, size=size)

                        if resampler is not None:

                            proces_label = resampler.Execute(label_image)

                        else:

                            proces_label = label_image



                        

                        # print(proces_image.GetSize(),proces_label.GetSize())

                        try:

                            assert proces_image.GetSize() == proces_label.GetSize()

                        except Exception as e:

                            failed_files.append(lf)

                            continue



                        label_output_path = os.path.join(output_dir, uid, TASK_VALUE, f"{lf_tissue}.nii.gz")



                        label_path_dict[lf_tissue] = label_output_path

                        util.save_nifti(proces_label, label_output_path, lf)

                        print(f"Saved Label Segment NIfTI file to {label_output_path}")

                        '''
                else:
                    label_flag=False
            except RuntimeError:
                failed_files.append(full_path)
                print(f"Failed to load DICOM images from {full_path}")
                continue
            
            '''

            meta.add_keyvalue('Image_id',meta_image_id)

            meta.add_keyvalue('Weeks',meta_weeks)

            meta.add_keyvalue('FVC',meta_fvc)

            meta.add_keyvalue('Percent',meta_percent)

            meta.add_keyvalue('Age',meta_age)

            meta.add_keyvalue('Sex',meta_sex)

            meta.add_keyvalue('Smoke_Status',meta_status)

            

            

            size_processed = list(proces_image.GetSize())



            meta_image_id=uid

            # meta.add_keyvalue('Image_id',meta_image_id)

            meta.add_keyvalue('Spacing_mm',min(spacing_info))

            meta.add_keyvalue('OriImg_path',full_path)

            meta.add_keyvalue('Size',size_processed)  # 这里用处理后的size -- YH Jachin

            meta.add_keyvalue('Modality',modality)

            meta.add_keyvalue('Dataset_name',study)

            '''
            roi='abdomen'
            if thorax_flag and gall_bladder_flag:
                roi='thorax-'+roi
            if thorax_flag and not gall_bladder_flag:
                roi='thorax'
            if pelvis_flag and gall_bladder_flag:
                roi=roi+"-pelvis"
            if pelvis_flag and not gall_bladder_flag:
                roi='pelvis'
            if lung_min>0 and lung_max/lung_min>3:
                label_dict[pid_dir]=[lung_max,lung_min]

            print(pid_dir,roi)
            #meta.add_keyvalue('ROI',roi)
            for ik in fj.keys():
                fi=fj[ik]
                jid=fi['Metadata']['Study_UID']
                max_length=fi['Spacing_mm']*max(fi['Size'])*0.001
                print(max_length,max_length>1.2)
                if jid==pid_dir:
                    if roi=='thorax-abdomen-pelvis' and max_length>1.2:
                        roi='whole-body'
                    fj[ik]['ROI']=roi
                    print(jid,max_length,roi)
                    break
                else:
                    continue
            
            '''  

            if label_flag:

                # print(label_path_dict.keys())

                meta.add_keyvalue('Task',TASK_VALUE)

                # meta.add_keyvalue('Label_tissue',list(label_path_dict.keys()))

                meta.add_keyvalue('Label_path',{TASK_VALUE:label_path_dict})



                # meta.add_keyvalue('Label_Dict',LABEL_DICT)



            meta.add_extra_keyvalue('Metadata',CIA_other_info)



            

            



            # Write the mapping to the JSON file on the fly

            with open(json_output_path, 'r+') as json_file:

                existing_mappings = json.load(json_file)

                existing_mappings[output_path] = meta.get_meta_data()

                json_file.seek(0)

                json.dump(existing_mappings, json_file, indent=4)

                json_file.truncate()

           '''
    else:
        print("No metadata.csv files found.")
    
  
    with open(json_output_path,'w') as fi:
        json.dump(fj,fi)
    print(f"The list has been written to {failed_files_path}")
    print(f"Saved NIfTI mappings to {json_output_path}")
    #print(label_dict)
if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Process NIIGZ files and save as NIfTI.")
    parser.add_argument("--target_path", type=str, help="Path to the target directory containing metadata files.", default="/home/data/Github/data/data_gen_def/DATASETS/AbdomenAtlas/uncompressed2")
    parser.add_argument("--output_dir", type=str, help="Directory to save the NIfTI files.", default="/home/data/Github/data/data_gen_def/DATASETS_processed/AbdomenAtlas_v2/")
    args = parser.parse_args()
    print(args.target_path, args.output_dir)
    main(args.target_path, args.output_dir)