A multimodal convolutional neural network for prostate cancer diagnosis using volumetric multiparametric MRI and clinical features.

目的：利用体积多参数磁共振成像(MpMRI)和临床特征，建立用于前列腺癌诊断的深度卷积神经网络(CNN)。动机：前列腺癌是英国男性最常见的癌症，2014-2035年期间发病率预计将上升12%。在英国，每10例前列腺癌病例中就有4-6例是在晚期确诊的[1]。通过结合解剖和功能信息，PROMIS(“前列腺MRI成像研究”)[2]的研究表明，前列腺癌的mpMRI可以从根本上改善前列腺癌的诊断，而且是非侵入性的。计算机辅助诊断技术可以用来从复杂的mpMRI数据和其他可用的临床数据中获得快速、经济有效和可重复的诊断[7]。在计算机视觉中，深层CNN正日益成为计算机辅助诊断的首选技术[4]。能够有效结合mpMRI数据的CNN架构是一个研究较少的问题，就像将成像数据与临床数据(如年龄、前列腺特异性抗原(PSA)水平和腺体体积)相结合一样。此外，基于CNN的方法将有利于PROMIS中强调的mpMRI的一些应用。例如，可以为患者分配0(健康)到1(不健康)之间的分数的CNN可以用于根据分数反映的异常程度对患者进行分类。背景/先前的工作：前列腺癌的计算机辅助诊断在过去十年中一直是一个活跃的研究领域。在文献[5]中，提出了一种新的基于手工特征的两阶段(体素分类/候选分类)计算机辅助检测流程。或者，基于CNN的方法“学习”最具区别性的特征，而不需要特征工程。对医学成像中300多项深度学习贡献的全面调查可以在[4]中找到。最近的一项最先进的贡献使用了共同训练的CNN来自动检测2D图像中的前列腺癌：T2加权(T2w)图像和表观扩散系数(ADC)图[8]。该项目的主要创新之处将是使用体积T2w图像、ADC图和动态对比度增强(DCE)图像。另一个新奇之处在于对DCE卷的处理。由于这些是在一系列时间步长上收集的，因此可能需要体素式统计方法来将DCE体积时间序列组合成单个体积。这项工作的另一个新方面将是对结合成像数据和临床数据的有效性的调查。最后一个新颖性将是将我们的CNN输出应用于分诊，据我们所知，这是一个迄今为止尚未探索的领域，可以产生重大的临床影响。

Aim: To build a deep convolutional neural network (CNN) for prostate cancer diagnosis that utilises volumetric multiparametric MRI (mpMRI) and clinical features.Motivation: Prostate cancer is the most common cancer amongst men in the UK, with incidence rates projected to rise by 12% in the period 2014-2035 [1]. 4-6 in 10 prostate cancer cases in the UK are diagnosed at a late stage [1]. By combining anatomical and functional information, studies like PROMIS ("Prostate MRI Imaging Study") [2] have shown that mpMRI of the prostate can radically improve prostate cancer diagnosis, non-invasively. Computer-aided diagnosis techniques can be used to obtain a fast, cost effective and reproducible diagnosis out of complex mpMRI data and other clinical data available [7]. In computer vision, deep CNNs are increasingly becoming the technique of choice for computer- aided diagnosis [4]. CNN architectures that can effectively combine mpMRI data is an understudied problem, as is combining imaging data with clinical data e.g. age, prostate specific antigen (PSA) level and gland volume. Furthermore, a CNN based approach would favour some of the applications of mpMRI highlighted in PROMIS. For example, a CNN that can assign a score to patients, between 0 (healthy) and 1 (unhealthy), may be used to triage patients according to the degree of abnormality reflected by the score.Background / Previous Work: Computer-aided diagnosis of prostate cancer has been an active area of research over the past decade. In [5], a novel two-stage (voxel classification / candidate classification) computer-aided detection workflow is developed based on handcrafted features. Alternatively, CNN based approaches "learn" the most discriminative features rather than requiring feature engineering. A comprehensive survey of over 300 contributions of deep learning in medical imaging can be found in [4]. A recent state of the art contribution uses a co-trained CNN for automated detection of prostate cancer in 2D images: T2-weighted (T2w) images and apparent diffusion coefficient (ADC) maps [8].Novelty: Most CNN approaches are only able to process 2D images whilst most medical data used in clinical practice consists of 3D volumes [6]. The main novelty of this project will be the use of volumetric T2w images, ADC maps and dynamic contrast enhanced (DCE) images. A further novelty will lie in the treatment of the DCE volumes. As these are collected over a series of timesteps, a voxel-wise statistical approach may be required to combine the DCE volumes time series into a single volume. Another novel aspect of this work will be an investigation of the effectiveness of combining imaging data with clinical data. A final novelty will be the application of our CNN output to triage, which to the best of our knowledge, is an area that is unexplored to date and can have a substantial clinical impact.