毕业生知识库

脑胶质瘤是成人中最常见的原发性脑肿瘤，这些脑胶质瘤患者中超过一半患 有致命的胶质母细胞瘤。神经外科手术是脑胶质瘤主要的治疗方式，其总体要求 是在保护神经功能的前提下实现肿瘤的最大限度切除。但由于脑胶质瘤具有浸 润生长的特点，神经外科医生凭借术中视诊与触诊很难在肿瘤边界处鉴别肿瘤 组织与正常组织，所以很容易导致肿瘤残留或患者功能区损伤。术中冰冻病理是 一种常规的诊断方法，可获得组织良恶性、肿瘤分级与 Ki-67 等信息，从而指导 手术切除与术后早期治疗策略制定。但它通常需要耗费很长时间（至少 20-30 分 钟），同时存在采样误差，而且需要专业的病理科医生配合对病理切片进行解读。 临床迫切需要一种新技术，可以自动对脑胶质瘤患者组织标本进行快速、准确的 术中诊断。

深度学习方法为脑胶质瘤术中快速诊断提供了新思路。通过结合深度学习与 近红外二区（NIR-II，near-infrared window II）荧光成像，使用深度神经网络自 动捕获荧光图像中的特征，将其与病理金标准建立关联，有望从中挖掘出更多肿 瘤相关信息，实现脑胶质瘤组织标本的术中快速、精确诊断。在此基础上，通过 使用网络架构搜索技术，针对性地优化网络结构，可以增强模型特征提取能力， 进一步提升术中诊断效果。总的来说，本文的主要工作和贡献概括如下：

（1） 针对脑胶质瘤组织术中快速、精确诊断的问题， 提出了一种基于 EfficientNet 迁移学习的术中脑胶质瘤诊断方法。该方法使用 ImageNet 预训练的 EfficientNet 网络在脑胶质瘤组织术中近红外二区荧光图像上进行迁移学习，通 过多种数据增强方法与加权损失函数缓解了训练过拟合与类别不平衡的问题， 实现了术中肿瘤组织和正常组织的快速鉴别，模型在测试集上的受试者工作特 征曲线下面积（AUC，area under receiver operating characteristic curve）可达 0.945。 同时该方法可以在术中实时提供肿瘤分级与 Ki-67 信息，AUC 分别达到 0.810 与 0.625。

（2） 针对手工设计专用网络结构过程较为复杂的问题，提出了一种卷积神经网络即插即用模块搜索方法 PP-NAS。该方法首先提出了一个新搜索空间 PPConv，可用于搜索即插即用模块。PPConv 可以很容易地整合到现有网络结构 中，提升模型多尺度特征提取能力。同时提出了一种基于单层优化的可微搜索算 法，通过使用额外的损失函数，减小了搜索阶段与评估阶段的优化间隙，提升了 搜索效果。搜索得到的网络在自然图像的图像分类、目标检测、语义分割等多个 任务上进行了验证，效果均优于对比网络模型。

（3） 针对术中脑胶质瘤组织的肿瘤鉴别、肿瘤分级、Ki-67 水平预测等问 题，在 PP-NAS 方法的基础上，根据脑胶质瘤术中组织荧光图像的特点，对搜索 空间进行了改进，并使用基于单层优化的搜索算法在脑胶质瘤数据集上开展了 搜索实验，进一步优化了模型结构，得到了更适合处理术中组织荧光图像的网络 结构，在肿瘤鉴别、肿瘤分级与 Ki-67 水平预测等任务上都取得了效果提升。

综上所述，针对脑胶质瘤术中诊断问题，本文提出了基于迁移学习与多尺度 即插即用模块搜索的术中脑胶质瘤诊断方法，在单中心脑胶质瘤数据集上进行 了验证，结果表明所提出的方法在术中肿瘤鉴别、肿瘤分级与 Ki-67 水平预测上 取得了良好的效果，有望在术中为医生提供帮助。

Glioma is the most common primary brain tumor in adults. Among these patients, more than half have glioblastoma, which is the most malignant glioma. Neurological surgery is the major treatment modality for glioma, in which maximal safe resection is required. However, because of the infiltrative nature of glioma, it is difficult for neurosurgeons to differentiate tumor tissues from normal tissues on the tumor boundary, resulting in tumor residual and damages to the eloquent areas of patients. Intraoperative frozen section analysis is a routine diagnostic method to obtain information on tumor and non-tumor, tumor grade and Ki-67 index, which can guide surgical resection and early postoperative treatment strategies. However, it is usually time-consuming (at least 20-30 min) and requires interpretation of intraoperative histologic images by experienced pathologists. Therefore, there is an urgent clinical need for a new technique that can provide rapid and accurate intraoperative diagnosis of tissue samples from glioma patients.

Deep learning provides a new way for rapid intraoperative diagnosis of glioma. By combining deep learning with the near-infrared window II (NIR-II) fluorescence imaging, we use deep neural networks to automatically capture features in fluorescence images and correlate images with pathologic diagnosis. It is expected that more tumor related information can be mined from fluorescence images to achieve rapid and accurate intraoperative diagnosis of glioma. Then, neural architecture search can be used to further optimize the network architecture, enhance the ability of feature extraction, and improve the performance of intraoperative diagnosis. Overall, the major work and contributions of this thesis are summarized as follows:

1. A transfer learning method based on EfficientNet for intraoperative diagnosis was proposed to solve the problem of rapid and accurate intraoperative diagnosis of brain glioma. ImageNet pre-trained EfficientNet was used to perform transfer learning on intraoperative fluorescence images of glioma. Then, heavy data augmentation and weighted loss function were applied during training to reduce over-fitting and class imbalance. With the proposed method, a rapid differentiation between tumor and non-tumor tissues was achieved and the area under the receiver operating characteristic curve (AUC) on the test set was 0.945. Furthermore, tumor grade and Ki-67 level could be predicted in real-time intraoperatively, and the AUC was 0.810 and 0.625, respectively.

2. Manual design of network architecture is complicated and sub-optimal. To mitigate this problem, a search method PP-NAS for searching plug-and-play modules in convolutional neural networks was proposed. PP-NAS includes a new search space for plug-and-play modules, which can be easily integrated into the existing networks to improve the ability of multi-scale feature extraction. A new search algorithm is also included, which reduces the optimization gap between the search and evaluation and improves the performance by using one-level optimization and extra loss functions. The searched network was superior to the compared networks in visual tasks of image classification, object detection and semantics segmentation.

3. To solve the problem of tumor differentiation, tumor grading and Ki-67 level prediction of glioma intraoperatively, based on the proposed PP-NAS, the search space was refined to fit the characteristics of intraoperative fluorescence images in glioma. Then, the search algorithm based on one-level optimization was used to carry out search experiments on the glioma dataset. As a result, a better network architecture for intraoperative fluorescence images was obtained, which improved the performance of tumor differentiation, tumor grading and Ki-67 level prediction.

In summary, methods based on transfer learning and multi-scale plug-and-play module search were proposed to solve the problem of intraoperative glioma diagnosis and evaluated on the single-center glioma dataset. The results showed that the proposed methods performed well in tumor differentiation, tumor grading and Ki-67 level prediction of glioma and were expected to provide help for neurosurgeons during surgery.