使用结直肠癌中的MMR标志物基于深度学习的MSI预测
Deep Learning based Prediction of MSI using MMR Markers in Colorectal Cancer
摘要：
结直肠癌的准确诊断和分子特征分析对于为患者规划最佳治疗方案至关重要。微卫星不稳定性（MSI）或错配修复（MMR）状态在适当的治疗选择中起着至关重要的作用，具有预后意义，并用于研究患者患有潜在遗传疾病（Lynch综合征）的可能性。NICE建议所有CRC患者都应进行MMR/MSI检测。免疫组织化学通常用于评估MMR状态，随后根据需要进行分子检测。这会产生大量额外费用，并需要额外资源。引入可以从目标图像预测MSI或MMR状态的自动化方法可以大大降低与MMR测试相关的成本。与先前关于MSI预测的研究（涉及使用粗标签训练CNN（MSI与微卫星稳定（MSS）不同，我们使用细粒度MMR标签进行训练。在本文中，我们提出了我们的工作，预测MSI状态，在两个阶段的过程中，使用一个单一的目标载玻片染色CK 8/18或H&E。首先，我们训练了一个多头卷积神经网络模型，其中每个头负责预测一个MMR蛋白质表达。为此，我们将MMR染色载玻片与目标载玻片配准作为预处理步骤。在第二阶段，从MMR预测图计算的统计特征用于最终MSI预测。我们的研究结果表明，MSI分类可以通过将细粒度的MMR标签相比，以前的方法，其中只有粗糙的标签被利用。
英文摘要：
The accurate diagnosis and molecular profiling of colorectal cancers are critical for planning the best treatment options for patients. Microsatellite instability (MSI) or mismatch repair (MMR) status plays a vital role in appropriate treatment selection, has prognostic implications and is used to investigate the possibility of patients having underlying genetic disorders (Lynch syndrome). NICE recommends that all CRC patients should be offered MMR/MSI testing. Immunohistochemistry is commonly used to assess MMR status with subsequent molecular testing performed as required. This incurs significant extra costs and requires additional resources. The introduction of automated methods that can predict MSI or MMR status from a target image could substantially reduce the cost associated with MMR testing. Unlike previous studies on MSI prediction involving training a CNN using coarse labels (MSI vs Microsatellite Stable (MSS)), we have utilised fine-grain MMR labels for training purposes. In this paper, we present our work on predicting MSI status in a two-stage process using a single target slide either stained with CK8/18 or H&E. First, we trained a multi-headed convolutional neural network model where each head was responsible for predicting one of the MMR protein expressions. To this end, we performed the registration of MMR stained slides to the target slide as a pre-processing step. In the second stage, statistical features computed from the MMR prediction maps were used for the final MSI prediction. Our results demonstrated that MSI classification can be improved by incorporating fine-grained MMR labels in comparison to the previous approaches in which only coarse labels were utilised.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
