Hybrid Intelligence for Trustable Diagnosis And Patient Management of Prostate Cancer (HIT-PIRADS)

项目总结/摘要 前列腺癌（PCa）是全世界男性最常见的癌症之一，估计有160万例病例和366万例患者。 每年死亡[1]。在美国，11%的男性在其一生中被诊断患有PCa，发病率通常随着年龄的增长而上升。 年龄[2]。前列腺成像报告和数据系统（PI-RADS）已成为使用 多参数磁共振成像（mp-MRI）。PI-RADS旨在标准化癌症分级的方法。然而，PI-RADS不使用临床和人口统计学患者信息，并且MR图像被定性或至多评估 半定量地导致危险癌症的检测不足和不重要癌症的过度检测。 该提案旨在开发人工智能（AI）算法，通过减少 评估变化并提供可靠的预测。我们的算法将使用不同的人口数据， 更好的评价体系。该新系统将输入mp-MRI、临床（直肠指检、PCa家族史）， 人口统计学（年龄、种族）和实验室（血清PSA）数据，以提供前列腺内病变的风险评分，以及 改善不同人群的患者管理。我们将开发的智能系统被称为混合智能 和可信赖（HIT）-PIRADS和本提案的具体目标有三个方面： 首先，我们将开发一个新的预处理框架，用于增强mp-MRI数据并最大限度地减少数据偏差。MRI 质量差异很大，这使得标准化非常困难。为了使MRI正常化，我们将纠正伪影， 不均匀性和噪声作为预处理步骤。接下来，数据集偏见，如种族的过度/不足代表性将被 因为偏见会导致扭曲和不准确的结果。我们将研究不平衡并量化数据中的不确定性 代表开发视觉偏差估计工具（ViBeT），以识别数据中的潜在偏差。二是 使用mp-MRI开发PCa的联合分割、检测和分类算法。前列腺定量和 PCa对于病变识别、风险分层、活检引导和手术/局灶性治疗的病变靶向至关重要。 我们将使用我们创新的基于胶囊的神经网络算法，并将其扩展到分析mp-MRI和非mp-MRI。成像数据。这一步将提高我们的算法对所有风险群体、种族和年龄的泛化能力。也会有 HIT-PIRADAS中的解释模块：我们将嵌入射线照相解释和视觉解释 进入基准的HIT-PIRADS。第三，我们将回顾性地评估和验证HIT-PIRADS的有效性， 和前瞻性。我们将在超过7000例患者的数据中证明HIT-PIRADS的有效性（3846例回顾性，3200例 前瞻性）。我们将严格评估变异来源，并将HIT-PIRADS标准化，以便在诊所中采用。 该项目的成果将是首个同类产品和易于使用的PCa检测推荐系统， 患者管理（HIT-PIRADS），以提供更准确、无偏倚、可重现的结果，减少PCa相关 发病率和死亡率。从长远来看，我们预计HIT-PIRADS将在诊所中广泛采用，并引发其他治疗 根据打击艾滋病毒/艾滋病综合行动计划制定预防战略。

Project Summary/Abstract Prostate Cancer (PCa) is among the most common cancers in men worldwide, with an estimated 1.6M cases and 366K deaths annually [1]. In the US, 11% of men are diagnosed with PCa over their lifetime, with incidence generally rising with age [2]. The Prostate Imaging Reporting and Data System (PI-RADS) has become a standard tool for diagnosing PCa using multi-parametric MR images (mp-MRI). PI-RADS aims to standardize the way to classify the cancer grades. However, PI-RADS does not use clinical and demographic patient information, and MR images are assessed qualitatively or at most semi-quantitatively causing under-detection of dangerous cancer and over-detection of insignificant cancer. This proposal is to develop artificial intelligence (AI) algorithms to improve the detection accuracy by reducing assessment variations and providing trustable predictions. Our algorithms will use diverse population data and eventually a far better evaluation system. This new system will input mp-MRI, clinical (digital rectal exam, PCa family history), demographic (age, race), and laboratory (serum PSA) data to provide risk scores for intraprostatic lesions, and improve patient management for diverse populations. The smart system we will develop is called Hybrid Intelligence and Trustable (HIT)-PIRADS and specific aims of this proposal are three-fold: First, we will develop a new pre-processing framework for enhancing mp-MRI data and minimizing data biases. MRI quality varies significantly, which makes standardization very difficult. To normalize MRI, we will correct artifacts, remove inhomogeneity and noise as the pre-processing step. Next, dataset bias, such as over/under-representation of race will be dealt with as biases cause skewed and inaccurate outcomes. We will examine imbalances and quantify uncertainties in data representation to develop a visual bias-estimation tool (ViBeT) to identify potential biases in the data. Second, we will develop joint segmentation, detection, and classification algorithms for PCa using mp-MRI. Quantification of prostate and PCa is essential for lesion identification, risk stratification, biopsy guidance, and lesion targeting for surgery/focal therapies. We will use our innovative capsule-based neural networks algorithms and extend its strength to analyze mp-MRI and nonimaging data. This step will improve generalization of our algorithms to all risk groups, races, and ages. There will be also an explanation module in the HIT-PIRADAS: we will embed both radiographical interpretations and visual explanations into the baseline HIT-PIRADS. Third, we will evaluate and validate the efficacy of the HIT-PIRADS both retrospectively and prospectively. We will prove the effectiveness of HIT-PIRADS in over 7000 patients’ data (3846 retrospective, 3200 prospective). We will rigorously evaluate sources of variations and standardize HIT-PIRADS for adoption in the clinics. The outcome of this project will be a first-of-its-kind and easy-to-use recommendation system for PCa detection and patient management (HIT-PIRADS) to provide more accurate, unbiased, reproducible results to reduce PCa related morbidity and mortality. In the long term, we expect HIT-PIRADS to be widely adopted in clinics and trigger other treatment & prevention strategies to be developed based on HIT-PIRADS.