Automatic segmentation of healthy tissues and tumours in patient brain images using 3D fully convolutional neural networks

据估计，目前有超过55，000名加拿大人患有脑肿瘤。神经胶质瘤是成人中最常见的原发性脑肿瘤，尽管研究取得了相当大的进展，但患者的预后仍然很差。高级别胶质瘤患者的平均预期寿命为两年或更短，而低级别胶质瘤患者的预期寿命为几年。在任何一种情况下，在治疗前后都采用神经影像学方案，以估计疾病进展、手术计划和效果。当前的临床协议涉及由放射科医师对患者图像的分析，其中采用基本的定性和定量度量，诸如肿瘤大小的手动测量，这是一个耗时、主观且可能不一致的过程。该项目的目标是开发强大、准确和全自动的组织分割技术，当应用于真实的、多模态、临床MRI时，可以识别健康和病变组织，并具有改善患者诊断、手术计划和随访的长期潜在受益。这包括开发新的机器学习（例如深度学习）技术，以在多通道患者MRI中准确检测和分割（1）肿瘤及其组成子结构（例如肿瘤核心、水肿）和（2）健康组织（例如白色物质）。尽管深度学习框架在计算机视觉的各种任务中取得了令人难以置信的成功，但它们对医学图像检测和分割的适应，特别是对病理结构的适应，仍处于起步阶段。这是由于在嘈杂的多模态3D图像的背景下提出了许多新的挑战，以及训练所需的大规模数据集的短缺。医学图像分析研究将受益于新的数学模型和分析工具的开发，这些模型和工具可能会改善患者护理和结果，包括节省时间和提高术前计划和术后随访的准确性。

It is estimated that over 55,000 Canadians are currently living with a brain tumour. Patients with gliomas, the most frequent primary brain tumour in adults, still have very poor prognosis despite considerable advances in research. High-grade glioma patients have a median life expectancy of two years or less, and low-grade gliomas come with a life expectancy of several years. In either case, neuroimaging protocols are employed before and after treatment in order to estimate disease progression, surgical planning and effect. Current clinical protocol involves analysis of the patient images by a radiologist, where rudimentary qualitative and quantitative metrics are employed, such as the manual measurements of tumour size, a process that is time-consuming, subjective and potentially inconsistent. The goals of this project are to develop robust, accurate and fully automatic tissue segmentation techniques that can identify both healthy and diseased tissues when applied to real, multimodal, clinical MRI, with the long-term potential benefit of improving patient diagnosis, surgical planning and follow-up. This includes the development of new machine learning (e.g. deep learning) techniques to accurately detect and segment (1) tumours into their constituent sub-structures (e.g. tumour core, edema) and (2) healthy tissues (e.g. white matter) in multi-channel patient MRI. Although deep learning frameworks have been incredibly successful at a wide variety of tasks in computer vision, their adaptation to medical image detection and segmentation, particularly of pathological structures, is still in its infancy. This is due to a multitude of new challenges presented in the context of noisy, multi-modal 3D images, and to a shortage of large-scale datasets required for training. Medical image analysis research would benefit from the development of new mathematical models and analytical tools that could potentially improve patient care and outcome, including the savings in time and the improvement in accuracy of pre-surgical planning and post-operative follow-up.