智能化肺结节辅助诊断中的人工智能技术研究

The 2015 global cancer statistics show that lung cancer accounts for approximately 13% of 14.1 million new cancer cases and 19.5% of cancer-related deaths each year. The 5-year survival for patients with an early diagnosis is approximately 54 %, as compared to 4 % if the diagnosis is made late when the patient has the stage IV disease. Hence, early diagnosis and treatment are the most effective means to improve lung cancer survival. On chest CT scans, a “spot” on the lung, less than 3 cm in diameter, is defined as a lung nodule, which can be benign or malignant. Malignant nodules may be primary lung tumors or metastases and so the classification of lung nodules is critical for best patient care..Radiologists typically read chest CT scans on a slice-by-slice basis, which is time-consuming, expensive and prone to operator bias and requires a high degree of skill and concentration. Computer-aided diagnosis (CAD) of lung nodule avoids many of these issues and has attracted a lot of research attention. Most solutions in the literature, however, are based on using hand-crafted image features to train a classifier, and hence are less capable of producing accurate diagnosis, due to the low soft-tissue contrast in CT images, complex morphology and location of nodules and limited number of chest CT images that can be used to construct the computational model. .This project will tackle the challenge of devising a much smarter, more sensitive and more reliable lung nodule CAD system based on the deep learning theory, which is the most profound breakthrough in recent artificial intelligence research. Our proposed system consists of four novel deep learning models, including a transferred attention-based deep semantic segmentation network (TA-DSSN) for nodule detection, a deep self-learning network based on adversarial example and weak model structure transfer (DSLN-AE/WMST) for nodule classification, a knowledge graph-based deep one-shot learning (KG-DOSL) model for nodule malignancy prediction, and a meta-learning-based deep life-long learning (ML-DL3) model for sustainable improvement of lung nodule diagnosis. .Millions of people, who are at high risk for lung cancer, rely for their well-being, and sometimes the delivery of life-saving treatment, on the quality and accessibility of early diagnosis. At stake is a sector in which we invest billions of dollars a year for that purpose. The achievements of this project will provide significantly more reliable and cost-effective lung nodule CAD technology, which will advance the artificial intelligence applied to medical imaging and be a major benefit for the health system and on the individual level, through opening up possibilities for creating niche domestic and export markets in advanced lung nodule CAD systems.

无论在中国，还是在全世界，肺癌都是发病人数和死亡人数最多的恶性癌症。由于恶性肺结节很有可能发展成肺癌，基于CT影像的肺结节计算机辅助诊断是实现肺癌早期诊断和肺癌筛查的一项关键技术。但是，由于CT影像具有较低的软组织对比度，且肺结节形态复杂，又有可能粘附于气管、血管和胸膜，难以准确检测和描述，再加上可供研究的数据较为有限，所以使得现有辅助诊断技术难以取得令人满意的效果。本项目拟以深度学习为代表的人工智能理论为基础，提出一套基于胸部CT图像的肺结节计算机辅助诊断技术，包括用于肺结节检测的基于迁移注意机制的深度语义分割网络、用于良恶性分类的基于对抗样本和弱模型结构迁移的深度自主学习模型、用于恶性度预测的基于知识图谱的小样本深度学习模型，以及旨在不断改善诊断性能的深度持续学习模型。本项目的成果不仅可以丰富人工智能在医学影像领域的理论与实践，还可以为肺癌早期诊断和肺癌筛查提供技术支撑，从而为建设“健康中国”做成贡献。

肺癌是全世界最常见和致死人数最多的癌症之一。研究表明，肺癌患者如果能够在病程早期被确诊并得到相应治疗，五年存活率可以达到60%；但四期肺癌患者的五年存活率仅有5%。数据显示，临床上肺癌患者的平均五年存活率仅19%。这是因为，超过70%的确诊患者都是晚期。由于肺结节在影像学表现密度增高的阴影，而恶性结节往往会发展为肺癌。因此，基于胸部CT影像的肺结节计算机辅助诊断是肺癌筛查中的一项关键技术，对于挽救肺癌患者的生命至关重要。但是，由于CT影像中肺结节与周围组织的对比度很低，且肺结节形态复杂，又大多粘附于气管、血管和胸膜上，难以准确检测、描述和诊断，再加上可供研究的数据非常有限，使得现有辅助诊断技术难以取得令人满意的效果。. 针对这些挑战，本项目重点研究了如何从CT影像中准确地检测肺结节，并对肺结节进行分割和恶性度预测，所取得的主要技术成果包括与领域知识相结合的深度协作学习技术、基于协同深度学习技术、非参数共享的半监督深度学习技术、基于样本对的自监督学习技术、基于注意力机制的不确定度学习技术、分割与分类相互引导的多任务弱监督学习技术、基于网络架构搜索技术的深度学习等，其中率先将与诊断相关的领域知识引入基于深度学习的肺结节良恶性分类，被加拿大工程院院士等国际著名学者誉为是“一项开创性的工作（a pioneer work）”；针对肺结节检测任务，在LUNA16数据集上取得了与目前排名第三的方法相当的性能；针对良恶性分类任务，在最大的数据集LIDC-IDRI上取得了目前最好的性能，同时，还进一步研究了标注偏差对肺结节恶性度预测的影响，以及可以有效使用不确定性标注信息的恶性度预测算法。. 项目期间，课题组累计获授权发明专利6项，另申请发明专利7项；在图像处理和机器学习领域著名国际学术期刊和会议发表论文68 篇，包括29 篇期刊论文（14篇论文发表在影响因子超过10的期刊上）和29 篇会议论文，其中发表在IEEE-TPAMI/TMI/TIP、MedIA、NeurIPS、CVPR、IJCAI、MICCAI等本领域顶级期刊/会议上的学术论文50余篇。本项目不仅开发一套基于CT影像的肺结节检测、分割及恶性度预测技术，为肺癌早筛提供了技术支撑，也丰富了深度学习的理论与实践，为其它医学影像分析问题的研究提供了有益借鉴。