肺结节检测数据

肺癌死亡率居高不下的原因除了其自身治疗困难、容易转移和复发等诸多因素外，没能早期发现、早期治疗也是很重要的原因。就肺癌而言，早期发现、治疗，是降低死亡率和提高生存质量的主要策略之一。因此，需要使用专业软件对肺癌进行早期筛查。 肺结节智能辅助诊断系统借助人工智能深度学习技术辅助放射科医生进行微小病灶早期筛查、精准定位、良恶性鉴别；自动匹配患者多次诊断影像，对影像中相同病灶进行跟踪对比，以列表和视图的方式呈现，让医生对患者病灶的发展（如倍增时间等）有一个直观的认识；同时根据诊断结果自动给出诊断意见以及生成清晰的、图文并茂的结构化报告。数据采集：1.CT扫描探测器大于等于16排；2.CT影像符合DICOM3.0标准；3.CT影像重建层厚3mm及以下；4.CT影像需为肺部的完整影像，且清晰无伪影。
数据脱敏：为保证患者隐私安全和患者利益，数据遵循相关法律法规和规范性文件的要求进行了脱敏处理。
数据预处理：数据清洗：按层厚、体位、是否存在缺陷、是否符合入排标准四个方面进行。1.层厚：要求层厚不大于3mm，若同一检查下有多个序列则取符合要求的最小层厚序列；2.是否存在缺陷：CT影像的扫描范围应包括肺尖与肋膈角之间的全部，且不能进行任何修改编辑，每个病例的影响应当保持连续完整，不得出现缺层、错层且图像清晰；3.统一图像方向：读取CT影像的方位信息，将图像三维像素矩阵转换为方向统一的三维矩阵。归一化：读取数据的像素信息，结合肺窗的窗宽、窗位进行归一化操作，减少干扰，增强肺部对比度。
数据标注：肺结节CT数据标注采取标注并审核的标注方式，每一例数据由标注医生背靠背分别进行标注。标注医生、审核医生要求在二级及以上医院任职，标注医生要求工作经验在5年以上，审核医生要求工作经验10年以上，职称为主任医师或更高。所有标注人员需经过培训。

The high mortality rate of lung cancer is attributable to multiple factors, including its inherent treatment difficulty, high risk of metastasis and recurrence, as well as the lack of early detection and treatment. Early detection and intervention for lung cancer are among the core strategies to reduce mortality and improve patients' quality of life. Therefore, professional software is required for the early screening of lung cancer.

The Pulmonary Nodule Intelligent Auxiliary Diagnosis System leverages artificial intelligence deep learning technology to assist radiologists in early screening, precise localization, and benign-malignant differentiation of small pulmonary lesions. It automatically matches multiple diagnostic images of the same patient, tracks and compares the same lesions across different imaging studies, and presents the comparison results in list and view formats, enabling clinicians to intuitively grasp the progression of the patient's lesions (e.g., doubling time). Additionally, it automatically generates diagnostic recommendations based on the examination results and produces clear, text-and-graphic structured reports.

### Data Collection
1. The CT scanner shall be equipped with ≥16 detector rows;
2. CT images must conform to the DICOM 3.0 standard;
3. The reconstructed slice thickness of CT images shall be ≤3 mm;
4. CT images must be complete lung scans with clear quality and no artifacts.

### Data Anonymization
To ensure patient privacy and safeguard patient interests, all data has been anonymized in accordance with relevant laws, regulations and normative documents.

### Data Preprocessing
#### Data Cleaning
The cleaning process is carried out from four dimensions: slice thickness, patient position, presence of imaging defects, and compliance with inclusion and exclusion criteria.
1. Slice Thickness: The slice thickness shall not exceed 3 mm. If multiple sequences are available for the same examination, the sequence with the minimum compliant slice thickness will be selected;
2. Imaging Defects: The CT scan range must cover the entire area between the lung apex and the costophrenic angle, without any manual modification or editing. Each case's images shall be continuous and complete, with no missing or misaligned slices, and the images must be of high clarity;
3. Unified Image Orientation: Read the orientation metadata of CT images, and convert the 3D pixel matrix into a uniformly oriented 3D matrix.

#### Normalization
Read the pixel value information of the data, and perform normalization operations combined with the window width and window level of the lung window to reduce image interference and enhance the contrast of pulmonary tissues.

### Data Annotation
The annotation of pulmonary nodule CT data follows a dual annotation and review workflow. Each case is annotated separately by two annotating physicians in a back-to-back manner. Both annotating and reviewing physicians must be employed in secondary or higher-level hospitals. Annotating physicians are required to have at least 5 years of relevant work experience, while reviewing physicians must have at least 10 years of work experience and hold the title of Chief Physician or higher. All annotation and review personnel must complete standardized training.