基于深度学习的脑垂体瘤MRI医学影像智能识别与诊断

Pituitary adenoma (PA) is one of the most common brain tumors. Early detection, precise diagnosis and personalized therapeutic plan of PA depend strongly on medical imaging. But since medical imaging interpretation is based on physicians’ recognition ability, there remain the problems as different judgment and interpretation of the PA among different doctors and different hospitals, and which in turn affect the therapeutic plan for PA patient. Medical imaging recognition, therefore, becomes the bottleneck of PA management. Deep learning in medical field (glioma, a kind of brain tumor, e.g.) shows its advantages by other studies, and our study also shows potential of deep learning on PA imaging recognition. However, the high cost of large data medical imaging labeling restricts the realization of artificial intelligence based on deep learning method, and the recognition rate and performance of image segmentation based on contour detection or random field method are unsatisfactory. Therefore, we intend to utilize deep learning on magnetic resonance imaging for PA, and to analyze the clinical data to achieve a standardized and personalized therapeutic plan according to big data. Our final aim is to establish the foundation of the intelligent clinical decision system based on deep learning, and to utilize deep learning in solving clinical problem. The implementation of this project will be of great theoretical significance and practical value.

垂体瘤的早期发现、精准诊断和个性化治疗方案制订依赖于医学影像，而以磁共振成像为主的医学影像对医生肉眼识别能力要求高，不同医生对垂体瘤的判读存在不确定性，进而使垂体瘤的治疗方案存在主观性、手术方案的不规范性。因此，对医学影像图像的判读成为垂体瘤治疗的瓶颈。深度学习在医学影像领域的初探显示了其优势，然而海量医疗影像的标注成本高，制约着基于深度学习在医学影像中的应用。本研究拟将深度学习应用于垂体瘤医学影像处理，首先通过半人工辅助标记进行边界精确标注，解决海量样本训练问题；后根据诊断报告和MRI图像训练垂体瘤分类网络，进而采用迁移学习训练垂体瘤语义分割网络，实现精准分割；在此基础上针对感兴趣区域进行纹理分析，并结合分析结果和患者生化指标，训练垂体瘤诊断和治疗网络，实现对垂体瘤的智能辅助诊断与治疗决策制定，达到通过深度学习的手段解决临床问题的目的。课题的实施将具有很好的理论探索意义和现实的使用价值。

Surgical planning may require preoperative evaluation of consistency in pituitary macroadenoma (PMA). Our study aimed to evaluate the diagnostic performance of a radiomic model based on multiparameter magnetic resonance imaging (mpMRI) in preoperative evaluation of tumor consistency in PMA. We applied automatic 3D (3D) segmentation to generate volume of interest (VOI) on T2WI and then co-register T1WI/T1CE on T2WI. A total of 388 radiomic features were extracted on each VOI of mpMRI. The top-discriminative features were identified using the minimum-redundancy maximum-relevance method and 0.632+ bootstrapping. The radiomics models based on each sequence and their combinations were established via the random forest (RF) and support vector machine (SVM), and independently evaluated for their ability in distinguishing PMA consistency. The radiomics model built on combined T1WI/T1CE/T2WI demonstrated the best performance to evaluate the PMA consistency and other sellar lesions.