Machine learning-based quantitative image, tissue, and clinical data analysis for lesion detection and characterization on prostate cancer imaging

前列腺癌通常通过全前列腺放射治疗或手术治疗，其副作用包括勃起功能障碍和尿失禁。然而，并非所有男性都需要这种侵入性治疗；一些前列腺癌生长非常缓慢，根本不需要治疗，而一些前列腺癌集中在一个部位，治疗只能针对肿瘤。前列腺癌的主要问题之一是如何为每个人确定正确的治疗方法。目前这是通过血液检查和穿刺活检来完成的。由于血液检查的准确性有限，而且小的活检针可能会遗漏肿瘤，医生需要在不完整的信息下选择治疗方法。为了解决这个问题，本研究提出了三维（3D）磁共振成像（MRI）和正电子发射断层扫描（PET）来完整地绘制前列腺。一种叫做前列腺特异性膜抗原（PSMA） PET成像的新形式，在帮助医生检测前列腺内癌症方面显示出巨大的希望。尽管MRI和PSMA PET是前列腺成像的最佳成像技术，但由于图像的复杂性，在MRI上看到前列腺肿瘤对医生来说仍然是非常困难的，而且目前还没有指导医生解释PSMA PET前列腺图像的指南。我们建议开发一种人工智能计算机系统，以帮助医生将复杂的图像转换为简单的3D癌症地图，从而指导活检目标并为每位患者选择正确的治疗方法。该系统将基于人工神经网络，这是一种模拟人类大脑视觉系统的软件程序。一种具有“深度学习”架构的新型人工神经网络在日常物体的机器视觉任务中显示出巨大的前景；例如，许多手机摄像头使用深度学习进行面部识别。我们将首次开发一种基于深度学习的人工智能系统，该系统将mpMRI和PSMA PET图像与医生将用于协助癌症检测的所有临床参数（例如血液检查结果）集成在一起。此外，我们将通过眼动追踪技术在医生和机器之间形成一个更强大的接口，从而进入医生的内心世界。眼睛注视的位置表明怀疑或不确定的区域，深度学习系统将利用这些眼睛注视的数据来提高对医生注意力所在的图像区域的评估。这将产生一种混合的人机视觉系统，利用人类和人工智能的综合优势，最终开发出一种廉价的软件工具，避免过度治疗前列腺癌的男性，因为过度治疗会损害生活质量，没有任何额外的好处，并在侵袭性前列腺癌仍可治愈的时候及早发现。

Prostate cancer is often treated by whole-prostate radiation treatment or surgery, with side effects including erectile dysfunction and urinary incontinence. However, not all men need such invasive treatment; some prostate cancers are sufficiently slow-growing to never need treatment, and some are concentrated in one spot and treatment can be targeted to the tumours only. One of the major problems in prostate cancer is how to determine the right treatment for each man. This is currently done using a blood test and a needle biopsy. Because the blood test has limited accuracy and the small biopsy needles may miss tumours, physicians need to select treatments using incomplete information. To solve this problem, this research proposes three-dimensional (3D) magnetic resonance imaging (MRI) and positron emission tomography (PET) to completely map the prostate. A new form of PET imaging, called prostate specific membrane antigen (PSMA) PET imaging, is showing tremendous promise in its ability to help physicians detect cancer within the prostate. Although MRI and PSMA PET are the best imaging technologies for prostate imaging, seeing prostate tumours on MRI can still be extremely for the physician difficult because of the complexity of the images, and there are no current guidelines for physicians to follow in interpreting PSMA PET images of the prostate. We propose the development of an artificially intelligent computer system to help doctors to translate the complex images into a simple 3D cancer map that will enable guiding of biopsies to targets and choosing the right treatment for each patient. This system will be based on artificial neural networks, which are software programs that mimic aspects of the visual systems in human brains. A new type of artificial neural network with a “deep learning” architecture has shown tremendous promise in machine vision tasks for everyday objects; for instance, many mobile phone cameras use deep learning for face recognition. We will, for the first time, develop a deep learning-based artificial intelligence system that will integrate mpMRI and PSMA PET images with all of the clinical parameters (e.g. blood test results) that the physician would use to assist in cancer detection. In addition, we will form a more robust interface between the physician and the machine by eye tracking technology to get a window into the physician's mind. Eye gaze locations suggest regions of suspicion or uncertainty, and the deep learning system will exploit these eye gaze data to sharpen its assessment of image regions where the physician's attention dwells. This will result in a hybrid human-machine vision system drawing on the combined strengths of human and artificial intelligence to finally develop an inexpensive software tool that will avoid overtreatment of prostate cancer in men for whom this would compromise quality life with no added benefit, and detect aggressive prostate cancer early while it is still curable.