基于深度学习的肺腺癌肿瘤突变负荷的预测

肿瘤突变负荷(TMB)与非小细胞肺癌(NSCLC)的免疫治疗疗效呈正相关,并且在近期的相关研究中,肿瘤突变负荷对靶向治疗及化疗的疗效也有一定的预测作用。因此,提出一种融合注意力机制的Inception深度学习模型(CAIM),用于对TCGA数据库中的非小型细胞肺癌中的肺腺癌的病理切片进行识别。首先,对数据样本进行切分,裁剪成小切片;然后送入到深度学习模型中,通过卷积学习图像特征,再与注意力机制结合进一步加强特征的提取;最后通过对小切片预测信息的整合,自动判别肺腺癌病理切片TMB值的高低。数据集由337张肺腺癌病理组织切片组成,其中高TMB值的数据271张,低TMB值的数据实验66张。结果表明,所提方法的性能平均曲线下面积(AUC)为0.82,明显高于图像分类方法残差网络(ResNet)的AUC值0.66。研究结果对临床实践中肿瘤突变负荷的检测和辅助诊断具有重要意义。

Prediction Tumor Mutation Burden of Lung Adenocarcinoma Based on Deep Learning

A number of existing medical studies have found out that tumor mutation burden (TMB) is positively correlated with the efficacy of non-small cell lung cancer (NSCLC) immunotherapy, and in the recent related studies, tumor mutation load also has a certain predictive effect on the efficacy of targeted therapy and chemotherapy. Based on above situations, this paper proposed an inception deep learning model CAIM (combine attention and inception-block module) to identify the pathological sections of lung adenocarcinoma in non-small cell lung cancer from the Cancer Genome Atlas (TCGA) dataset. First, by segmenting the data samples, cutting them into small slices, and then sending them to the deep learning model, learning image features through convolution, and then combining with the attention mechanism to further strengthen the feature extraction. Finally, through the integration of the prediction information of the small slices, the TMB value of the pathological tiles of lung adenocarcinoma (LUAD) could be automatically determined. The data set consisted of 337 LUAD pathological tissue sections, including 271 data with high TMB value and 66 data experiments with low TMB value. Experimental results showed that the averaged area under the curve (AUC) of the proposed method was 0.82, which significantly better than the AUC value of 0.66 for the residual network of image classification method and was of great significance for the detection of tumor mutation burden and auxiliary diagnosis in clinical practice.