Deep learning for renal tumor characterization

肾肿瘤特征的深度学习

• 批准号: 10116348
• 负责人: Harrison Bai
• 金额: $ 4.16万
• 依托单位: RHODE ISLAND HOSPITAL
• 依托单位国家: 美国
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-03-01 至 2021-06-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10116348
• 关键词: 3-Dimensional; Ablation; Algorithms; Architecture; Area; Benign; Biopsy; Characteristics; Classification; Clear cell renal cell carcinoma; Clinic; Clinical; Clinical Data; Clinical Management; Computer software; Data; Detection; Diagnosis; Dropout; Ensure; Excision; Future; Growth; Histology; Image; Image Analysis; Indolent; Institution; Intervention; Interventional radiology; Kidney; Kidney Neoplasms; Learning; Lesion; Literature; Machine Learning; Magnetic Resonance Imaging; Malignant - descriptor; Medical Imaging; Metastatic Neoplasm to the Kidney; Modeling; Nephrectomy; Nephrons; Neural Network Simulation; Normal tissue morphology; Oncology; Operative Surgical Procedures; Outcome; Output; Patient imaging; Patients; Performance; Process; Renal Cell Carcinoma; Renal Mass; Research Project Grants; Selection for Treatments; Signal Transduction; Techniques; Therapeutic; Training; Triage; Update; Urologist; Weight; base; cancer imaging; clinical application; clinical decision-making; cohort; deep learning; deep neural network; design; imaging modality; improved; interest; learning network; novel; outcome prediction; predictive modeling; radiologist; radiomics; random forest; treatment response; tumor

Our long-term objective is to develop deep learning techniques capable of predicting characteristics and treatment response or response to surveillance to assist clinical decision- making in renal tumors that are potential candidates for ablation therapy, biopsy, active surveillance or surgical resection. An increasing number of renal tumors are being diagnosed, due in part to incidental detection from the increased use of cross-sectional imaging. Although partial nephrectomy is still considered the primary treatment for small renal masses, percutaneous ablation is increasingly performed as a therapeutic, nephron-sparing approach. One challenge for interventional radiologists and urologists who manage these patients is selection for therapy, since the average rate of progression is slow for small renal tumors and metastasis rarely occurs. A technique that could distinguish indolent tumors from those will progress based on data from the imaging methods used to detect and delineate renal masses would enable early triage to observation versus invasive treatment. Deep learning, a type of machine learning technique which takes raw images as input, and applies many layers of transformations to calculate an output signal, has already led to breakthroughs in other areas of image recognition, and is increasingly used for medical image analysis. However, its application in the field of interventional radiology is currently limited. Furthermore, no study in the literature has applied deep learning to kidney lesion segmentation and characteristics/outcome prediction. In this project, we propose to develop novel deep learning architectures based on routine MR imaging that allow for accurate renal mass segmentation and prediction of characteristics and outcome in renal tumors. Using data from four independent cohorts, we will use our deep learning architectures to predict (1) benign versus malignant histology (2) growth rate in stage 1a renal cell carcinoma (3) SSIGN score in clear cell renal cell carcinoma and (4) clinical endpoints. We will integrate segmentation and classification into one net that suitable for clinical application. In addition, we will compare results with those of experts and traditional machine learning approaches.

我们的长期目标是开发能够预测 特征和治疗反应或对监测的反应，以协助临床决策- 使肾肿瘤成为消融治疗、活检、活性 监视或手术切除。越来越多的肾肿瘤被诊断出来， 部分原因是由于横截面成像的使用增加而引起的偶然检测。虽然 肾部分切除术仍然被认为是小肾肿块的主要治疗方法， 经皮消融越来越多地作为治疗性的保留肾单位的方法来执行。 介入放射科医生和泌尿科医生管理这些患者的一个挑战是 选择治疗，因为小肾肿瘤的平均进展速度缓慢， 转移很少发生。一种可以区分惰性肿瘤和 基于用于检测和描绘肾脏肿块的成像方法数据的进展 将使早期分诊观察与侵入性治疗。深度学习，一种 机器学习技术，以原始图像作为输入，并应用多层 变换来计算输出信号，已经导致了其他领域的突破， 图像识别，并且越来越多地用于医学图像分析。然而，其应用 目前在介入放射学领域的应用是有限的。此外，文献中没有研究 将深度学习应用于肾脏病变分割和特征/结果预测。 在这个项目中，我们建议开发基于常规MR的新型深度学习架构， 成像允许准确的肾肿块分割和特征预测， 肾肿瘤的结果。使用来自四个独立队列的数据，我们将使用我们的深度 学习体系结构，以预测（1）良性与恶性组织学（2）阶段生长率 1a肾细胞癌（3）透明细胞肾细胞癌的SSIGN评分和（4）临床 端点。我们将分割和分类集成到一个适合临床的网络中， 应用程序.此外，我们将与专家和传统机器的结果进行比较 学习方法。