基于欠采样的影像组学机器学习模型术前预测子宫肌瘤高强度聚焦超声消融效果

打开文本图片集

Abstract：ObjectiveToimprovetheaccuracyofmachinelearningmodelsforpreoperativepredictionofhigh-intensityfocused ultrasound （HIFU）ablationeficacyforuterine fibroidsbycorecting classimbalance insmallsample datasetsusing undersamplingmethods.MethodsClinicalandimaging datawerecolectedfrom140patientswithuterinefibroids undergoingHFUtreatmentatFoshanWomenandChildrenHospital，including104withhighablationratesand36withlow ablationrates.Radiomic features were extracted from MRIT-weighted images （TWI）of the patients，and machine learning modelswereconstructedtopredictHIFUtreatmentoutcomes.Fourmachinelearningalgorithms，includingk-Nearest Neighbors （KN）andomForest （RF），upportVectorchine（V），andultilayerPerceptron （MP），ereoupledith 7undersamplingmethods，namelyRandomUndersampling （RUS），RepeatedEditedNearest Neighbors（RENN），AlkNearestNeighbors（AllN），NeighborhoodCleaningRule-3（NM），CondensedNearestNeighbor（CNN），Neigboro CleaningRule（NCR）andstanceHardnessTreshold（I），forandinglassimbalanceinthedatasets.The28iction modelswereevaluated using 5-fold cross-validationforareasunderthereceiveroperating characteristiccurve （AUC）， acuracyrecallandscificityesultseestombinationsofdersampligmethodsadmacineaingmodelsC RF，NM-，NN-KNN，andN-PadAUsof0.772 95% CI： 0.566-0.942）， 0.797 （ 95% CI： 0.600-0.950）， 0.822 （95% CI： 0.635-0.964）， and 0.822 （ 95% CI： 0.632-0.960），respectively. The AUCs of themachine learning models significantly increased aftercoupling with undersampling methods， withthe MLPmodel showing the most pronounced improvement.Thereall rates of the4combinedmodelsalsoimprovedsignificantly（by0.389forCNN-RF，0836forNM-VM，0.532forCNN-KNN， and0.372 for NM-MLP）. Conclusion The use of undersampling methodscan efectively correct classimbalance in small sample datasets to improve the accuracyof machinelearning models for predicting the eficacy of HIFU ablation foruterine fibroids.

Keywords：uterinefibroid;magneticesonanceimaging;highintensityfocusedultrasound;machinelearing;prediction;class imbalance; radiomics; undersampling

子宫肌瘤是生殖系统最常见的肿瘤之一，在女性生命历程中，约 70% 的白人女性和 80% 的非裔美国女性会受此疾病困扰。（剩余18417字）