Children's Cancer Therapy Development Institute, Beaverton, OR, 97005, USA.
Omics Data Automation, Beaverton, OR, 97005, USA.
Mod Pathol. 2022 Sep;35(9):1193-1203. doi: 10.1038/s41379-022-01075-x. Epub 2022 Apr 21.
Correctly diagnosing a rare childhood cancer such as sarcoma can be critical to assigning the correct treatment regimen. With a finite number of pathologists worldwide specializing in pediatric/young adult sarcoma histopathology, access to expert differential diagnosis early in case assessment is limited for many global regions. The lack of highly-trained sarcoma pathologists is especially pronounced in low to middle-income countries, where pathology expertise may be limited despite a similar rate of sarcoma incidence. To address this issue in part, we developed a deep learning convolutional neural network (CNN)-based differential diagnosis system to act as a pre-pathologist screening tool that quantifies diagnosis likelihood amongst trained soft-tissue sarcoma subtypes based on whole histopathology tissue slides. The CNN model is trained on a cohort of 424 centrally-reviewed histopathology tissue slides of alveolar rhabdomyosarcoma, embryonal rhabdomyosarcoma and clear-cell sarcoma tumors, all initially diagnosed at the originating institution and subsequently validated by central review. This CNN model was able to accurately classify the withheld testing cohort with resulting receiver operating characteristic (ROC) area under curve (AUC) values above 0.889 for all tested sarcoma subtypes. We subsequently used the CNN model to classify an externally-sourced cohort of human alveolar and embryonal rhabdomyosarcoma samples and a cohort of 318 histopathology tissue sections from genetically engineered mouse models of rhabdomyosarcoma. Finally, we investigated the overall robustness of the trained CNN model with respect to histopathological variations such as anaplasia, and classification outcomes on histopathology slides from untrained disease models. Overall positive results from our validation studies coupled with the limited worldwide availability of sarcoma pathology expertise suggests the potential of machine learning to assist local pathologists in quickly narrowing the differential diagnosis of sarcoma subtype in children, adolescents, and young adults.
正确诊断肉瘤等罕见的儿童癌症对于确定正确的治疗方案至关重要。由于全球范围内专门从事儿科/青年肉瘤组织病理学的病理学家数量有限，因此对于许多全球区域来说，在病例评估早期获得专家的鉴别诊断机会有限。在中低收入国家，缺乏高技能的肉瘤病理学家的问题尤为突出，尽管肉瘤发病率相似，但病理学专业知识可能有限。为了解决这个问题，我们开发了一种基于深度学习卷积神经网络（CNN）的鉴别诊断系统，作为一种预病理学家筛选工具，根据经过培训的软组织肉瘤亚型的整个组织病理学切片，量化诊断的可能性。该 CNN 模型是在一个由 424 个经过中央审查的肺泡横纹肌肉瘤、胚胎性横纹肌肉瘤和透明细胞肉瘤肿瘤的组织学幻灯片组成的队列上进行训练的，这些肿瘤最初在原发机构诊断，并随后由中央审查进行验证。该 CNN 模型能够准确地对保留的测试队列进行分类，所有测试的肉瘤亚型的接收者操作特征（ROC）曲线下面积（AUC）值均高于 0.889。随后，我们使用 CNN 模型对一组来自人类肺泡和胚胎性横纹肌肉瘤的外部来源样本以及一组来自横纹肌肉瘤基因工程小鼠模型的 318 个组织学切片进行分类。最后，我们研究了训练有素的 CNN 模型在组织病理学变化方面的整体稳健性，例如间变，以及对未经训练的疾病模型的组织学幻灯片的分类结果。我们的验证研究的总体积极结果，加上全球肉瘤病理学专业知识的有限可用性，表明机器学习有可能帮助当地病理学家快速缩小儿童、青少年和年轻成人肉瘤亚型的鉴别诊断范围。
