第一部分：

研究背景：

经鼻蝶入路手术（TSS）是库欣病的首选治疗手段，其术后即刻缓解与否对远期预后和医患沟通至关重要。目前还没有基于多种机器学习方法的术后即刻缓解的预测模型。本研究开发并且评价了九种基于机器学习的模型，可在术前预测库欣病患者经鼻蝶手术后的即刻缓解情况。

方法：

根据入组标准，在2000年2月到2019年9月之间，共有1045名库欣病患者被纳入研究，被随机分入训练集（n=836）和测试集(n=209)。我们总共使用了九种机器学习算法，将11个术前结构化特征作为输入构建机器学习模型。在训练集上得到的模型在测试集中进行验证以评估模型的表现。我们使用受试者工作曲线（ROC）的曲线下面积（AUC）来评估模型的效果。

结果：

所有患者中，库欣病术后即刻缓解率为73.3%（766/1045）。在Logistic回归分析的单因素分析中，非首次手术（P<0.001），术前MRI上肿瘤侵袭海绵窦（P<0.001），大腺瘤（P<0.001），术前血浆ACTH的水平（P=0.008）以及病程（P=0.01）与患者的术后即刻缓解密切相关。各个机器学习模型的AUC值分布在0.664到0.743之间。其中表现最好的模型由堆叠式集成学习算法训练而来，此模型包含了四个术前特征：是否首次手术，术前MRI上是否有肿瘤侵袭海绵窦，肿瘤是否为大腺瘤以及术前血浆ACTH水平。

结论：

我们可以把患者的术前临床特征作为输入，并使用多种机器学习算法构建术后即刻缓解的预测模型，从而有利于医患沟通和手术决策的制定。

第二部分：

研究背景：

术前制定垂体腺瘤手术方案需要观察肿瘤的形态、大小及毗邻结构，垂体腺瘤的影像组学也需要高精度的肿瘤分割。目前垂体腺瘤影像组学中的肿瘤分割都是手动分割，消耗大量人力及时间成本，而且不同医生之间对感兴趣区的标注差异较大。本研究旨在开发一种自动分割的模型，用于肿瘤感兴趣区的分割以及提取与肿瘤侵袭性相关的特征，并且在临床任务中评估其有效性。本研究旨在通过影像组学的方法，术前预测垂体腺瘤的质地。

方法：

筛选北京协和医院2000年2月到2019年9月所有符合纳入标准的患者作为研究对象。我们使用了一种门控U网络（gated-shaped U-net）的深度卷积神经网络将鞍区结构分割为八个部分。根据鞍区磁共振分割的结果选取了五个特征，包括肿瘤直径，体积，视神经交叉的高度，Knosp分级以及肿瘤对颈内动脉的包绕程度，最终通过对垂体腺瘤质地的术前预测来判断此自动分割方法的有效性。

结果：

    共计163例垂体腺瘤患者被纳入到第一组当中并且被随机分配到训练集（131例）和测试集（32例）中。50例肢端肥大症患者被纳入到第二组中。在MRI的关键层面上垂体腺瘤模型预测的Dice值为0.940。本研究提出的方法在预测五个侵袭性相关的MRI特征时准确性超过80%。我们开发的自动分割模型在临床模型和影像组学上的AUC值分别达到了0.840和0.920。

结论：

我们开发的自动分割模型可以自动分割鞍区结构并且以较高的准确性提取重要特征。同时在预测垂体腺瘤质地这个临床问题上，我们的模型表现良好，这验证了我们模型临床应用的可行性。因此我们的自动分割模型也可以推广到其它的临床任务上，比如预后预测，放疗反应性等。

第三部分：

研究背景：

    去中心化的联邦学习技术可以用于多中心的临床研究，它可以最大化地利用多中心数据且不需要分享数据。我们旨在提出一个联邦学习流程，指导基于机器学习技术的多中心临床研究，并且使联邦学习模型的表现达到与使用集中化数据训练的模型一样的效果。

方法：

    我们开发了联邦学习流程，以指导机器学习在临床分类任务中的应用。本研究纳入了598例来自北京协和医院和120例来源于宣武医院的肢端肥大症患者。来自于协和医院的患者又被分为5部分。我们选取9个临床特征作为输入，并通过四个算法构建模型。我们用受试者工作曲线的曲线下面积（AUC）来评估模型的表现。

结果：

    基于联邦学习流程训练出来的LR模型比大部分单中心数据训练出来的LR模型表现要好（P<0.05）；基于联邦学习流程训练出来的DNN, SVM以及GBDT模型比所有单中心训练出来的模型都表现得要好（P<0.05）。基于联邦学习流程训练出来的LR, DNN以及SVM模型与集中化数据训练出来的模型效果无统计学差异（P>0.05）。

结论：

    我们开发了一个联邦学习流程，在进行多中心机器学习模型的训练时，无需分享各个中心的数据。此流程不仅可以用于垂体腺瘤的临床研究，也可以用于神经外科其他病种的临床研究，甚至用于其他专业的临床科研工作。

section I:

Background:

Transsphenoidal surgery is the first option for the treatment of Cushing’s disease. The immediate remission of Cushing’s disease is of significant importance to long-term prognosis. It is also very important in doctor-patient communication. There are no established accurate models that use machine learning (ML) methods to preoperatively predict immediate remission after transsphenoidal surgery (TSS) in patients diagnosed with histology-positive Cushing’s disease (CD). Our current study aims to devise and assess an ML-based model to preoperatively predict immediate remission after TSS in patients with CD.

Methods:

According to the inclusion and exclusion criteria, a total of 1,045 participants with CD who received TSS at Peking Union Medical College Hospital in a 20-year period (between February 2000 and September 2019) were enrolled in the present study. In total nine ML classifiers were applied to construct models for the preoperative prediction of immediate remission with preoperative factors. The area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate the performance of the models. The performance of each ML-based model was evaluated in terms of AUC.

Results:

The overall immediate remission rate was 73.3% (766/1045). First operation (P<0.001), cavernous sinus invasion on preoperative MRI(p<0.001), tumour size (P <0.001), preoperative ACTH (P=0.008), and disease duration (P=0.010) were significantly related to immediate remission on logistic univariate analysis. The AUCs of the models ranged between 0.664 and 0.743. The highest AUC, i.e., the best performance, was 0.743, which was achieved by stacking ensemble method with four factors: first operation, cavernous sinus invasion on preoperative MRI, tumor size and preoperative ACTH.

Conclusions:

We developed a readily available ML-based model for the preoperative prediction of immediate remission in patients with CD.

Section II:

Background:

The resection plan of pituitary adenoma (PA) needs preoperative observation of the sellar region. Radiomics prediction requires high-quality segmentations. Manual delineation is time-consuming and subject to rater variability. This work aims to create an automated segmentation method for the sellar region, several tools to extract invasiveness-related features, and evaluate their clinical usefulness by predicting the tumor consistency.

Methods:

Patients included were diagnosed with pituitary adenoma at Peking Union Medical College Hospital. A deep convolutional neural network, called gated-shaped U-net (GSU-Net), was created to automatically segment the sellar region into 8 classes. Five magnetic resonance imaging (MRI) features were extracted from the segmentation results, including tumor diameters, volume, optic chiasma height, Knosp grading system, and degree of internal carotid artery contact. The clinical usefulness of the proposed methods was evaluated by the diagnostic accuracy of the tumor consistency.

Results:

A total of 163 patients with confirmed pituitary adenoma were included as the first group and were randomly divided into a training data set and test data set (131 and 32 patients, respectively). Fifty patients with confirmed acromegaly were included as the second group. The Dice coefficient of pituitary adenoma in important image slices was 0.940. The proposed methods achieved accuracies of more than 80% for the prediction of 5 invasive-related MRI features. Methods derived from the automatic segmentation showed better performance than original methods and achieved areas under the curve of 0.840 and 0.920 for clinical models and radiomics models, respectively.

Conclusion:

The proposed methods could automatically segment the sellar region and extract features with high accuracy. The outstanding performance of the prediction of the tumor consistency indicates the methods’ clinical usefulness for supporting neurosurgeons in judging patients’ conditions, predicting prognosis, and other downstream tasks during the preoperative period.

Section III:

Background: Decentralized federated learning (DFL) may serve as a useful framework for machine learning (ML) tasks in multicentered studies, maximizing the use of clinical data without data sharing. We aim to propose the first workflow of DFL for ML tasks in multicentered studies, which can be as powerful as those using centralized data.

Methods: A DFL workflow was developed with four steps: registration, local computation, model update, and inspection. A total of 598 participants with acromegaly from PUMCH, and 120 participants from XWH were enrolled. The cohort from PUMCH was further split into five centers. Nine clinical features were incorporated into ML-based models trained based on four algorithms: LR, GBDT, SVM, and DNN. The area under the curve (AUC) of receiver operating characteristic curves was used to evaluate the performance of the models.

Results: Models trained based on DFL workflow performed better than most models in LR (P<0.05), all models in DNN, SVM, and GBDT (P<0.05). Models trained on DFL workflow performed as powerful as models trained on centralized data in LR, DNN, and SVM (P>0.05).

Conclusions: We demonstrate that the DFL workflow without data sharing should be a more appropriate method in ML tasks in multicentered studies. And the DFL workflow should be further exploited in clinical researches in other departments and it can encourage and facilitate multicentered studies.