Increasing Nodule Detection in Lung Cancer by Non-Conscious Detection of "Missed" Nodules and Machine Learning

通过无意识检测“遗漏”结节和机器学习来增加肺癌结节检测

• 批准号: 10626108
• 负责人: Gregory James DiGirolamo
• 金额: $ 44.3万
• 依托单位: UNIV OF MASSACHUSETTS MED SCH WORCESTER
• 依托单位国家: 美国
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-05-23 至 2027-01-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10626108
• 关键词: Anatomy; Arousal; Behavioral; Biological Markers; Brain; Cancer Detection; Cessation of life; Chest; Classification; Complex; Conscious; Data; Detection; Diagnosis; Diagnostic; Diagnostic Errors; Early Diagnosis; Eye; Feedback; Image; Learning Module; Left; Location; Lung; Lung nodule; Machine Learning; Malignant Neoplasms; Malignant neoplasm of lung; Measures; Medical Students; Mission; Modality; Modeling; Mus; Nodule; Normal tissue morphology; Outcome; Patients; Physiological; Principal Investigator; Process; Public Health; Pupil; Radiology Specialty; Reading; Research; Specific qualifier value; Speed; Survival Rate; Testing; Time; Training; X-Ray Computed Tomography; biomarker identification; biomedical imaging; cancer survival; clinical practice; detection limit; experience; improved; indexing; innovation; machine learning model; novel; programs; radiologist; random forest; visual search; visual tracking

Lung cancer has a 5-year survival rate of 21% and more than 80% of all new patients are diagnosed at an advanced stage. Finding small lung nodules representing the early stages of lung cancer are critical, but diagnostic error can be as high as 50% in harder-to-detect lung nodules. Nodule detection is the outcome of a difficult search task which employs both conscious and non-conscious brain processes that radiologists spend years of training enhancing. Currently, detection is constrained to those processes that become conscious. A critical need is characterizing and utilizing these non-conscious processes to improve nodule detection beyond conscious detection limits. This proposed RO1 will use an innovative new paradigm to isolate non-conscious processes during lung nodule searches in CT images. Using eye-tracking, the project will show clear and reliable biomarkers of non-conscious detection for “missed” nodules in the absence of any conscious detection or consideration of the nodule. These biomarkers will be used to train Machine Learning (ML) to detect “missed” nodules. The innovation of this application is capitalizing on the full expertise of the radiologist by utilizing these biomarkers of non-conscious detection to develop ML to read the radiologist and not the image; disrupting the status quo of ML in radiology, and creating ML that can detect “missed” nodules. The central hypothesis will be tested in three Specific Aims: 1: To evaluate the extent to which identified biomarkers of non-conscious detection of missed nodules can be used to train and refine ML models to increase nodule detection; 2: To quantify the extent that feedback of the locations that ML models indicate are missed nodules can increase nodule detection; and 3: To specify the extent that trained ML models can generalize to a novel set of radiologists on a novel set of chest CTs to detect missed lung nodules and increase nodule detection. These Aims will be carried out by testing radiologist on lung nodule searches in CT images using high-speed eye- tracking and our innovative paradigm that allow us to isolate non-conscious processes during misses and demonstrate that non-conscious processes are successfully detecting the “missed” nodules. ML models will be trained on these non-conscious biomarkers to detect “missed” lung nodules. The ML models will provide significant feedback to the radiologists to reduce the number of nodules missed by the limits of conscious detection. The proposed research is significant because it is expected to provide strong scientific justification for the use of non-conscious processes in diagnostic visual search, and to create ML models capable of detecting otherwise “missed” lung nodules; hence, changing clinical practice, reducing nodule misses, improving early detection, and increasing lung cancer's 5-year survival rate.

肺癌的5年生存率为21%，超过80%的新患者在接受治疗时被诊断为肺癌。 晚期发现代表肺癌早期阶段的小肺结节至关重要，但 在较难检测的肺结节中，诊断错误可高达50%。诺阿检测是一个 一项困难的搜索任务，它使用了放射科医生花费在有意识和无意识大脑过程中的信息。 多年的训练，不断提高。目前，检测仅限于那些成为有意识的过程。一 关键的需求是表征和利用这些无意识的过程，以改善结节检测， 意识检测极限这个提议的RO 1将使用一个创新的新范式来隔离无意识 在CT图像中的肺结节搜索期间处理。使用眼动追踪，该项目将显示清晰， 在没有任何有意识检测的情况下，无意识检测“遗漏”结节的可靠生物标志物 或考虑结节。这些生物标志物将用于训练机器学习（ML）来检测 “漏诊”结节。该应用程序的创新是充分利用放射科医生的专业知识， 利用这些无意识检测的生物标志物来开发ML以读取放射科医生而不是图像; 打破放射学中ML的现状，并创建可以检测“遗漏”结节的ML。中央 将在三个具体目的中检验假设：1：评价在何种程度上鉴定的生物标志物， 遗漏结节的无意识检测可用于训练和细化ML模型，以增加结节数量。 检测; 2：量化ML模型指示的位置反馈是遗漏结节的程度 可以增加结节检测;以及3：指定训练的ML模型可以推广到新集合的程度 放射科医生使用一套新的胸部CT来检测漏诊的肺结节并提高结节检测率。这些 目标将由测试放射科医生使用高速眼在CT图像中搜索肺结节- 跟踪和我们的创新范式，使我们能够隔离无意识的过程中错过， 证明无意识的过程成功地检测到了“错过”的结节。ML模型将 接受这些无意识生物标志物的训练，以检测“遗漏”的肺结节。ML模型将提供 向放射科医生提供重要反馈，以减少因意识障碍而遗漏的结节数量。 侦测这项拟议中的研究意义重大，因为它有望提供强有力的科学依据 在诊断视觉搜索中使用无意识过程，并创建能够 检测否则“遗漏的”肺结节;因此，改变临床实践，减少结节遗漏， 改善早期发现，提高肺癌的5年生存率。