Cyst-X: Interpretable Deep Learning Based Risk Stratification of Pancreatic Cystic Tumors

项目摘要 该项目的总体目标是开发一种新的诊断工具，称为Cyst-X，用于准确检测和 癌前胰腺囊肿的特征和通过精确的决策改善患者预后 (手术切除或监视)。胰腺癌是所有癌症中最致命的癌症，因为它的致命性很差。 预后差，缺乏早期检测方法。与其他常见癌症不同的是，前驱病变很好 已知(结肠息肉-结肠癌，导管原位癌(DCIS)-乳腺癌)，胰腺癌先兆 (囊肿病)人们知之甚少。早期诊断胰腺癌可能会降低死亡率和 这种致命疾病的发病率。早期诊断胰腺癌的一个主要方法是 在胰腺癌前肿瘤(囊肿性)转化为浸润性癌之前将其作为靶点。曾经的囊肿症 通过核磁共振成像(MRI)等放射成像检测到的，应确定其特征 关于它们的恶性潜在性。低风险的囊肿仍然是无害的；因此，患者应该留在 监视程序。另一方面，高危囊肿可能发展为侵袭性癌症，因此， 如果可能，患者应接受手术切除。尽管如此，风险分层的国际指南 胰腺囊肿的特征严重不足(确定低风险与高风险特征的准确率为55%-76% 囊性肿瘤的风险，而MRI检测囊性肿瘤的准确率只有40%-50%)。加在一起，这些关键障碍 提示迫切需要改进胰腺囊性肿瘤的特征。基于我们的 初步结果，支持基于图像的诊断决策工具的开发，我们假设 我们提出的Cyst-X将对胰腺囊肿的特征进行更高的诊断准确率，并提供 与目前的指导方针相比，更好地管理患者。对于这一总体假设，我们将 首先使用强大的深度学习方法(特别是深层胶囊网络)自动检测和 从多序列MRI扫描中分割胰腺和胰腺囊肿(目标1)。接下来，我们将创建一个 可解释的基于图像的胰腺囊肿诊断模型(目标2)。准确 对于这样的诊断模型，定性是必要的；然而，也将重点放在 机器生成的诊断模型的可解释性。辨别性特征的视觉解释将 帮助放射科医生在患者管理中获得更高的决策率。在目标3中，我们将验证建议的 Cyst-X框架的多中心研究。总共将从三个项目中收集1200个多序列MRI扫描 参与的临床中心(梅奥诊所、哥伦比亚大学医学中心、伊拉斯谟医学中心)。 将进行综合评估，以测试Cyst-X的有效性和普适性。所有评估都将是 根据国际指南和活组织检查证实的地面事实而制定。我们提议的研究具有广泛的 具体地说，从长远来看，它将影响胰腺癌的早期诊断和临床 提高胰腺癌生存率的决策。

Project Summary The overall goal of this project is to develop a new diagnostic tool, called Cyst-X, for accurate detection and characterization of pre-cancerous pancreatic cysts and improve patient outcome through precise decisions (surgical resection or surveillance). Pancreatic cancer is the most fatal cancer among all cancers due to its poor prognosis and lack of early detection methods. Unlike other common cancers where precursor lesions are well known (colon polyps-colon cancer, ductal carcinoma in situ (DCIS)-breast cancer), pancreas cancer precursors (cysts) are poorly understood. Diagnosing pancreatic cancer at earlier stages may decrease mortality and morbidity rates of this lethal disease. One major approach for diagnosing pancreatic cancer at earlier stages is to target pancreatic precancerous pancreatic neoplasms (cysts) before they turn into invasive cancer. Once cysts are detected with radiology imaging such as magnetic resonance imaging (MRI), they should be characterized with respect to their malignant potential. Low-risk cysts remain harmless; hence, patients should remain under surveillance program. On the other hand, high-risk cysts can progress into an aggressive cancer, therefore, patients should undergo surgical resection if possible. Despite this, international guidelines for risk stratification of pancreatic cysts are woefully deficient (55-76% accuracy for determining characteristics of low-risk vs high risk cystic tumors, while only 40-50% accuracy detecting cysts with MRI). Combined, these critical barriers indicate that there is an urgent need for improving characterization of pancreatic cystic tumors. Based on our preliminary results, which support the development of an image-based diagnostic decision tool, we hypothesize that our proposed Cyst-X will produce higher diagnostic accuracy for characterizing pancreatic cysts and provide better patient management compared to the current guidelines. Towards this overarching hypothesis, we will first use powerful deep learning methods (specifically deep capsule networks) for automatically detecting and segmenting the pancreas and pancreatic cysts from multi-sequence MRI scans (Aim 1). Next, we will create an interpretable image-based diagnosis model for characterizing pancreatic cysts (Aim 2). Accurate characterization is necessary for such a diagnostic model; however, emphasis will also be placed on interpretability of the machine generated diagnostic model. Visual explanation of the discriminative features will help radiologists obtain higher decision rates in patient management. In Aim 3, we will validate the proposed Cyst-X framework in a multi-center study. A total of 1200 multi-sequence MRI scans will be collected from three participating clinical centers (Mayo Clinic, Columbia University Medical Center, Erasmus Medical Center). Comprehensive evaluations will be made to test the validity and generalizability of Cyst-X. All evaluations will be made with respect to the international guidelines and biopsy proven ground truths. Our proposed study has wide implications: specifically, in the long term, it will influence early diagnosis of pancreatic cancer and clinical decision making to improve survival rates of pancreatic cancer.