Development of a Multimodal Deep Learning Model for the Generation of Cancer Probability Maps and Imaging Biomarkers for Prostate Cancer using Multiparametric MRI

项目摘要/摘要 背景：前列腺癌是最常见的新诊断癌症，也是死亡率第二高的癌症。 美国男性罹患癌症。该病的发病率和死亡率之间存在着很大的差异。 因此，开发筛查工具来识别前列腺癌并确定它是侵袭性的还是惰性的 是一个令人相当感兴趣的领域。目前的方法依赖于血清生物标记物和后续活检的使用。 用于筛查。然而，关于筛查的适当方法仍有很大的争议。的目标是 这项提议是开发：1)前列腺癌的新的成像生物标记物(即，“特征”)；以及2)a) 侵袭性前列腺癌的新预测模型。这些工具将实现更多 MP-MRI在未来前列腺癌筛查中的有效应用，从而使未来在前列腺癌筛查中的 筛查的敏感性和特异性，降低了误诊率和漏诊率。 目的1：实现一种用于临床前列腺癌MP-MRI序列的深度学习算法，创建一个癌的可能性。 预测活组织检查结果的能力图。 目标2：创建一个多模式框架，将已发现的成像特征与临床数据相结合 从医疗记录(例如，年龄、风险因素、病史、生物标志物)中预测存在和 前列腺癌的侵袭性。 方法：在目标1中，将在包含以下内容的临床数据集上训练深度卷积神经网络(CNN 从前列腺癌患者的前列腺癌切除前MP-MRI序列中提取的斑块，使用His-His- 全前列腺根治术标本的病理组织学分析。这方面的创新之处 目的是开发一种可以同时从三种不同成像序列类型中学习的CNN， 使用补丁进行数据增强，以及MP-MRI序列和前列腺切除术的正确比对 用于机器学习的样本。这一目标的工作结果将是创建一种算法，用于生成- 从MP-MRI数据中评估成像生物标记物(特征)和癌症概率图。在目标2中，多式联运 学习框架，将MP-MRI序列数据与临床参数相结合，以预测 会出现侵袭性前列腺癌。这一目标的创新将是- 开发可集成来自多个医疗模式(成像、血清、病史等)的信息的框架 为了产生对侵袭性前列腺癌存在的高置信度预测，而不使用 侵入性测试。 长期目标：开发一种新的侵袭性前列腺存在的预测模型 前列腺癌的MP-MRI数据，将使未来能够更好地利用这些数据来早期检测前列腺癌 前列腺癌。

Project Summary/Abstract Background: Prostatic adenocarcinoma is the most common newly diagnosed cancer and second deadliest cancer in American men. There is a large discrepancy between the incidence of the disease and its mortality rate. Thus, the development of screening tools to identify prostate cancer and determine if it is aggressive or indolent is an area of considerable interest. Current methods rely on the use of serum biomarkers and follow-up biopsies for screening. However, there is substantial debate as to the appropriate methodology for screening. The goal of this proposal is the development of: 1) new imaging biomarkers (i.e., “features”) for prostate cancer; and 2) a novel predictive model for the presence of aggressive prostatic adenocarcinoma. These tools will enable more effective use of mp-MRI in prostate cancer screening in the future and thus enable a future improvement in the sensitivity and specificity of screening, reducing the rates of overdiagnosis and underdiagnosis. Aim 1: To implement a deep learning algorithm for clinical prostate mp-MRI sequences, creating a cancer prob- ability map that is predictive of biopsy results. Aim 2: To create a multimodal framework that will combine discovered imaging features with clinical data points from the medical record (e.g., age, risk factors, medical history, biomarkers) to predict the presence and aggressiveness of prostatic adenocarcinoma. Methods: In Aim 1, a deep convolutional neural network (CNN) will be trained on a clinical dataset comprised of patches extracted from pre-prostatectomy mp-MRI sequences from patients with prostate cancer, using his- topathology analysis of whole-mount radical prostatectomy specimens as ground truth. The innovations in this aim will be the development of a CNN that can simultaneously learn from three different imaging sequence types, the use of patches for data augmentation, and the proper alignment of mp-MRI sequences and prostatectomy specimens for machine learning. The result of the work of this aim will be the creation of an algorithm for gen- erating imaging biomarkers (features) and cancer probability maps from mp-MRI data. In Aim 2, a multimodal learning framework that will integrate mp-MRI sequence data with clinical parameters in order to predict the presence of aggressive prostatic adenocarcinoma will be developed. The innovation in this aim will be the devel- opment of a framework that can integrate information from multiple modalities (imaging, serum, history, etc.) in order to generate a high confidence prediction of the presence of aggressive prostate cancer without the use of invasive testing. Long-term Objective: The development of a novel predictive model for the presence of aggressive prostatic adenocarcinoma in prostate mp-MRI data that will enable better future use of this data for the early detection of prostate cancer.