研究背景：鞍区囊性及囊实性病变是一类重要的鞍区病变类型，其种类多样，这些病变在临床实践中发现比例较鞍区实质病变少，但流行病学研究显示这一比例可能被低估。研究目的：本研究希望通过术前影像学检查对鞍区囊性及囊实性病变病理类型加以区分并对不同病变的内分泌影响模式差异进行探究。本研究将以鞍区囊性及囊实性病变的钆对比增强磁共振影像为材料，使用影像组学及人工神经网络等方法，对鞍区常见囊性病变的病理类型进行影像鉴别；通过对其内分泌实验室检验结果的分析，探究常见的鞍区囊性病变对各垂体内分泌轴功能的影响。研究方法：本研究筛选出过去15年间就诊于北京协和医院的诊断明确、具有完整的术前使用钆对比剂增强磁共振原始影像数据的、患有鞍区含囊性成分的疾病的患者。图像经过配准、神经网络自动感兴趣区域分割、人工验证分割区域后，提取三维、二维最大层面的影像组学特征，使用统计及传统机器学习算法、全连接神经网络对病变的影像特征进行两种病变间的分类，评估各种算法模型表现，同时应用卷积神经网络对原始磁共振影像二维最大层面进行直接识别评估效果。本研究同时对拥有内分泌实验室检验数据的病例的生长激素轴、甲状腺轴、性腺轴损伤及其病变尺寸、类型间关系进行了探究。对于病例数过少的垂体脓肿、淋巴细胞性垂体炎、转移癌，本研究进行了描述。研究结果：本研究纳入399例患者，对垂体瘤卒中、囊性垂体腺瘤、Rathke裂囊 肿、囊性颅咽管瘤进行两种病变间的影像鉴别。基于支持向量机的模型对囊性颅咽管瘤与Rathke裂囊肿和囊性垂体腺瘤鉴别度最佳，曲线下面积可达 0.8522。囊 性垂体腺瘤和垂体瘤卒中鉴别困难，在不同特征组、不同模型下均表现接近不可分类，这可能和其病理基础间的关联性有关系。逻辑斯蒂回归提示病变类型并非生长激素轴、甲状腺轴、性腺轴损伤与否的关联自变量，近似的体积-损伤分析提示随着垂体卒中病变体积增大，发生损伤的内分泌轴依次为性腺、生长激素、甲状腺。研究结论：通过影像组学、机器学习及神经网络方法能够对垂体瘤卒中、囊性垂体腺瘤、Rathke裂囊肿、囊性颅咽管瘤的增强磁共振成像的大部分影像进行鉴 别，垂体瘤卒中和囊性垂体腺瘤不可鉴别，不同病变类型对生长激素轴、甲状腺轴、性腺轴的损伤无显著差异。

Background: Cystic lesions are important parts of lesions of the sellar area. There are various types of lesions, including pituitary apoplexy, Rathke's cleft cyst, cystic craniopharyngioma, etc. These cystic lesions are found to be less common than solid lesions in clinical practice, but epidemiological research shows that their proportion may be underestimated.Purpose: This study aimed to distinguish the pathological types of cystic lesions in sellar area through preoperative magnetic resonance imaging examinations and to explore the influence modes of different lesions on endocrine function of the pituitary glands. This study utilized gadolinium-enhanced magnetic resonance imaging of cystic lesions to identify the common pathological types of cystic lesions in the sellar area with radiomics methods and deep-learning approaches. Through the analysis of the endocrine laboratory examinations, the influence of common sellar cystic lesions on the functions of the pituitary endocrine axis was investigated.Methods: This study screened patients who had a clear diagnosis of sellar disease with cystic components at Peking Union Medical College Hospital in the past 15 years, who had complete preoperative scans of gadolinium-enhanced magnetic resonance imaging. After the series were registered, an artificial neural network for automatic segmentation was applied to the images to annotate the regions of interest, and the segmented regions were manually confirmed later. Both the three-dimensional image and the two- dimensional at the largest layer were used to extract radiomics features. The extracted features were analyzed using statistics, traditional machine learning algorithms, and a fully connected neural network, to classify in pairs to evaluate the performance of these algorithm models. Convolutional neural networks were used to directly identify and evaluate the effect of the largest layer of two-dimensional images of the original magnetic resonance series. We also explored the relationship between the growth hormone axis, thyroid axis, and gonadal axis damage and the size and type of lesions with endocrine laboratory examination data. We further described clinical patterns of pituitary abscess, lymphocytic hypophysitis, and metastatic cancer, which had too few cases to be analyzed with quantitative methods. Results: In this study, 399 patients were enrolled. Paired imaging differentiations were performed on pituitary apoplexy, cystic pituitary adenoma, Rathke's cleft cyst, and cystic craniopharyngioma. The model based on support vector machine could simply distinguish cystic craniopharyngioma from Rathke's cleft cyst or cystic pituitary adenoma, and the area under the curve can be 0.8522. Cystic pituitary adenomas and pituitary apoplexy were difficult to distinguish, and they were almost unclassifiable with any algorithms on any feature sets, which may be related to their pathological basis.Logistic regression suggested that the type of lesion was not an independent variable of the damage of growth hormone axis, thyroid axis, or gonadal axis. The analysis of the relationship between the damage and the volume of the lesion with an estimate method suggested that as the size of the pituitary apoplexy increased, the endocrine axis of injury happened with the order of the gonadal, the growth hormones, and the thyroid sequentially. Conclusion: Through radiomics, machine learning and artificial neural network approaches, most of the pituitary apoplexy, cystic pituitary adenoma, Rathke's cleft cyst, and cystic craniopharyngioma can be distinguished with contrast-enhanced magnetic resonance imaging. Pituitary apoplexy and cystic pituitary adenoma cannot be distinguished with these methods. There is no significant difference between the pathology type and the damage of the growth hormone axis, thyroid axis, and gonadal axis.