Detecting pre-cancerous lesions from high resolution prostate MRI

描述(由申请人提供)： 成功地治疗已经转移的癌症比在癌变过程的早期治疗癌症或癌前状态要困难得多。前列腺癌，如果及早发现，五年存活率为100%--与许多其他类型的癌症相比，这是一个令人惊讶的积极统计数据。因此，通过筛查及早发现和定位前列腺癌是至关重要的。最近，我们一直在开发计算机辅助诊断(CAD)工具，用于从高分辨率磁共振成像(MRI)中检测恶性和癌前病变。由于人们普遍认为癌前病变会转化为癌症，这样的系统将有助于识别和监测前列腺癌的高风险患者，并启动早期靶向治疗，以逆转肿瘤过程。该项目广泛的长期目标是及早发现前列腺癌前病变和恶性病变，这在以下方面具有极其重要的意义：(1)监测前列腺癌的高风险患者；(2)早期有针对性地治疗前列腺癌前病变和恶性病变；(3)检测新的组织学类型，这可能对了解疾病过程具有重要意义。这项工作的首要目标是早期发现癌前病变和恶性病变，并可能通过CAD在高分辨率体外磁共振成像上识别新的组织类型。拟开展的工作共包括3个具体目标和9项任务。由于众所周知，癌前病变经常与前列腺癌共存，在这项研究中，我们建议只包括被诊断为前列腺癌并计划接受前列腺癌切除术的患者。在目标1下，将获得总共20个匿名患者数据集，包括3个特斯拉(T)体外磁共振扫描，以及根治性前列腺切除术后的整体组织切片。组织学数据的纳入将允许通过H&E染色和手动分割精确确定癌前病变的存在和程度。目的1还将通过将整个组织切片与相应的体外MRI图像配准来确定体外癌前病变(地面真相)的空间范围。为了检测癌前病变的存在和空间范围，我们采用了双管齐下的方法，使用监督和非监督分类技术。首先，在目标2下，我们开发和评估了一种有监督的CAD方法，通过在体外MRI研究中显式模拟HGPIN的纹理属性来检测癌前病变。在目标3中，我们从一个有监督的CAD模型开始，在体外的MRI上区分前列腺癌和良性病变，然后应用一种无监督的非线性降维方法来检测那些具有介于良性和恶性之间的特征的新的组织类型。AIM 3将提供(I)检测癌前病变的辅助方法，从而有助于评估在Aim 2中开发的有监督的CAD模型的有效性，以及(Ii)有助于潜在地发现新的组织学分类，从而促进我们对癌症进展的了解。目标2、3中提出的方法的有效性将根据组织学的基本事实进行评估。这个项目将是罗格斯大学和宾夕法尼亚大学(UPenn)研究人员之间的合作。目标1(数据生成)将在宾夕法尼亚大学进行，而目标2(肿瘤地面真相生成)和目标3(CAD模型)将在罗格斯大学进行。

DESCRIPTION (provided by applicant): Successfully treating cancer that has metastasized is considerably more difficult than treating the cancer or precancerous state early in the process of carcinogenesis. Prostate cancer, if caught early, has a 100 percent, five-year survival rate - a surprisingly positive statistic compared to many other types of cancer. For this reason, early detection and localization of prostate cancer through screening is critical. Of late we have been developing computer-aided diagnosis (CAD) tools for detecting malignant and pre-malignant lesions from high resolution Magnetic Resonance (MR) imaging (MRI). Since pre-malignant lesions are widely believed to transform into carcinoma, such a system will help identify and monitor patients with a high risk of prostate cancer and initiate early targeted treatment for regression of the neoplastic process. The broad long term goal of this project is early detection of pre-malignant and malignant prostate lesions, which is extremely significant in (1) Monitoring patients with a high risk of developing prostatic adenocarcinoma, (2) Early targeted treatment for regression of pre-malignant and malignant lesions, and (3) Detection of new histological tissue classes which may be significant in understanding disease processes. The overarching goal of this work is early detection of pre-malignant and malignant lesions and possible identification of new histological tissue classes on high-resolution ex vivo MR imagery via CAD. The proposed work comprises a total of 3 specific aims and 9 tasks. Since it is known that pre-malignant lesions frequently coexist with prostate carcinoma, in this study we propose to only include patients who have been diagnosed with prostate cancer and have been scheduled for a prostatectomy. Under Aim 1 a total of 20 anonymised patient data sets comprising 3 Tesla (T) ex vivo MRI scans with accompanying whole mount histological sections after radical prostatectomy will be obtained. The inclusion of histological data will allow for precise determination of presence and extent of pre-malignant lesions via H&E staining and manual segmentation. Aim 1 will also involve determination of spatial extent of pre-malignant lesions (ground truth) ex vivo by registering the whole mount histological sections with the corresponding ex vivo MRI images. To detect presence and spatial extent of pre-malignant lesions we adopt a two pronged approach using a supervised and unsupervised classification technique. First, under Aim 2 we develop and evaluate a supervised CAD method for detecting pre-cancerous lesions by explicitly modeling textural attributes of HGPIN on ex vivo MRI studies. Under Aim 3 we begin with a supervised CAD model to distinguish cancerous from benign prostate lesions on ex vivo MRI and then apply an unsupervised non-linear dimensionality reduction method to detect new histological tissue classes as those that have characteristics which are intermediate between benign and malignant. Aim 3 will provide (i) a secondary method of detecting pre- cancerous lesions and thus useful in evaluating efficacy of the supervised CAD model developed in Aim 2 and (ii) aid in potential discovery of new histological classes which could facilitate our understanding of cancer progression. The efficacy of the methods proposed under Aims 2, 3 will be evaluated against ground truth derived from histology. This project will be a collaboration between investigators at Rutgers University and the University of Pennsylvania (UPENN). Aim 1 (Data Generation) will be carried out at UPENN while Aims 2 (Tumor ground truth generation) and 3 (CAD model) will be done at Rutgers.