MLICP-CNN:基于CNN与ICP的多标记胸片置信诊断模型

针对胸片的多标记预测集缺少可校准性的缺陷,提出一种基于卷积神经网络(Convolutional Neural Networks,CNN)与归纳一致性预测器(Inductive Conformal Prediction,ICP)的多标记胸片置信诊断模型MLICP-CNN。该模型将学习数据划分为训练集和校准集,通过使用CNN从训练集中学习出规则D。基于规则D和校准集使用算法随机性对被测数据进行置信预测,即为每个被测数据提供附带置信度的多标记预测集。在对Chest X-ray14胸片数据集的实验结果表明,该模型在临床常用的95%置信度下,模型准确率为95%,体现了置信度评估的恰好可校准性。在CNN架构为Resenet50并采用LS-MLICP为奇异值映射函数下,模型性能最好,其确定预测率为96.43%,理想预测率为92.31%。另外,CNN架构对预测效率的影响程度远远小于奇异值映射函数。

MLICP-CNN: MULTI-LABEL CHEST X-RAY CONFIDENCE DIAGNOSIS MODEL BASED ON CNN AND ICP

To address the absence of calibrated confidence evaluation of multi-label prediction for chest x-ray,we proposed a multi-label chest X-ray confidence diagnosis model based on CNN and ICP,named MLICP-CNN. Our model divided the learning data into training set and calibration set,and a rule D was learned from the training set through CNN. Based on rule D and calibration set,we used the randomness of the algorithm to predict the confidence of the measured data,that is,to provide a multi-label prediction set with confidence for each measured data. The experimental results on the chest X-ray14 set demonstrate that the accuracy rate of MLICP-CNN is exactly 95% under the confidence levels of 95% in common clinical,revealing the exactly validity of confidence evaluation. In addition,when using Resenet50 as the component of CNN framework and adopting LS-MLICP as a nonconformity measure,our model gains the best performance with the certain prediction of 96.43% and favorite prediction of 92.31%. The influence of CNN framework on prediction efficiency is significantly less than that of the nonconformity measure.