MRI纹理分析联合机器学习模型对前列腺腺泡癌Gleason分级的预测价值

Predictive value of MRI texture analysis combined with machine learning model for Gleason grading of prostate acinus cancer

目的探讨利用前列腺腺泡癌MRI纹理分析对前列腺腺泡癌Gleason评分(GS)高危组(GS≥4+3分)与低危组(GS≤3+4分)进行预测的价值。方法选择前列腺根治全切术后病理诊断为前列腺腺泡癌的148例患者,年龄43~90岁,平均年龄61.18岁;其中高危(GS≥4+3分)80例,低危(GS≤3+4分)68例;血清前列腺特异抗原(PSA)为0.70~161.92 ng/mL,PSA中位数7.16 ng/mL;前列腺体积(PV)为13.42~174.81 cm^(3),平均PV 45.32 cm^(3);前列腺病灶体积为0.26~122.70 cm^(3),病灶体积中位数2.71 cm^(3);病灶位置外周带(PZ)69例,交界地带13例,移行带(TZ)66例。患者均行高分辨率横轴面T2加权成像(T;WI)和横轴扩散加权成像(DWI)扫描(b值=1500 s/mm^(2))。利用ITKSNAP软件勾画三维(3D)病灶,从MRI图像中提取T;WI和表观扩散系数(ADC)图像的影像组学纹理特征,依次采用Mann-Whitney U检验,最小冗余和最大相关算法(MRMR)筛选可预测前列腺癌高低危组的纹理特征子集,然后用单变量分析所选纹理特征与两组之间的关联性,并对单个所选纹理特征的诊断效能进行评估。随后基于筛选出的相对重要特征,采用随机森林(RF)分类器构建预测模型,对预测模型进行受试者工作特性曲线分析,评估模型的效能,并计算模型的灵敏度、特异度、准确度、阳性预测值和阴性预测值。最后采用留组交叉验证法(LGOCV)评估模型的可靠性。结果148例待预测病例中Gleason评分为:3+3分42例,3+4分26例,4+3分21例,4+4分27例,4+5分19例,5+4分8例,5+5分5例。高危组病灶PZ 37例,交界地带13例,TZ 30例;病灶体积1.55~13.30 cm^(3),病灶体积中位数4.00 cm^(3)。低危组病灶PZ 32例,交界地带0例,TZ 36例;病灶体积0.95~3.95 cm^(3),病灶体积中位数1.60 cm^(3)。筛选出10个相对重要的纹理特征在高低危组间具有显著差异,其中T2_wavelet_HHL_glszm_GrayLevelNonUniformity、T2_log_sigma_2_0_mm_3D_glszm_GrayLevelNonUniformityNormalized、ADC_wavelet_HLH_glszm_LargeAreaHighGrayLevelEmphasis预测高低危组前列腺腺泡癌的曲线下面积(AUC)分别达到了0.73、0.72、0.71。最后基于筛选出的10个纹理特征联合构建RF分类器预测模型,模型的准确度、灵敏度、特异度及AUC分别为0.70、0.63、0.77、0.76,阳性预测值和阴性预测值分别为0.75、0.65。结论基于双参数MRI纹理特征分析联合机器学习模型对高危和低危前列腺腺泡癌的预测具有临床应用价值。

Objective To investigate value of magnetic resonance imaging(MRI)texture analysis in prediction of Gleason score(GS)for high-risk group(GS≥4+3 scores)and low-risk group(GS≤3+4 scores)of prostate acinus cancer.Methods A total of 148 patients with prostatic acinus carcinoma confirmed by pathology after radical prostatectomy and total prostatectomy were enrolled,which aged 43-90 years old with mean age of 61.18 years old;of which,80 in high-risk group(GS≥4+3 scores)and 68 in low-risk group(GS≤3+4 scores).The serum prostate specific antigen(PSA)was 0.70-161.92 ng/mL with median PSA of 7.16 ng/mL;prostate volume(PV)was 13.42-174.81 cm^(3)with mean PV of 45.32 cm^(3);prostate lesion volume was 0.26-122.70 cm^(3)with median of 2.71 cm^(3);69 cases were at peripheral zone(PZ),13 at junction zone(JZ)and 66 at transitional zone(TZ).All of them were performed high-resolution transverse T2-weighted imaging(T;WI)and transverse diffusionweighted imaging(DWI)scans(b=1500 s/mm^(2)).The ITKSNAP software was used to outline 3 D lesions,and radiomics texture features of T;WI and apparent diffusion coefficient(ADC)images were extracted from MRI images.Mann-Whitney U test,the min-redundancy and max-relevance(MRMR)was used to screen and predict feature subsets of prostate cancer high-risk and low-risk group,the univariate was used to analyze correlation between texture features and 2 groups.The diagnostic efficiency of selected texture features was evaluated.Then,based on selected relatively important features,the random forests(RF)method was used to construct prediction model.The receiver operating characteristic(ROC)curve analysis was performed on prediction model to evaluate effectiveness of model,the sensitivity,specificity,and accuracy,positive predictive value and negative predictive value of model were calculated.Finally,leave group out cross-validation(LGOCV)was used to evaluate reliability of model.Results The GS of 148 cases was 3+3 scores in 42 cases,3+4 scores in 26,4+3 scores in 21,4+4 scores in 27,4+5 scores in 19,5+4 scores in 8 and 5+5 scores in 5.In high-risk group,there were 37 cases at PZ,13 at JZ and 30 at TZ;lesion volume was 1.55-13.30 cm^(3)with median lesion volume of 4.00 cm^(3).In low-risk group,there were32 cases at PZ,0 at JZ and 36 at TZ;lesion volume was 0.95-3.95 cm^(3)with median lesion volume of 1.60 cm^(3).Ten relatively important texture features were selected and showed significant differences between high-risk and low-risk group,in which the area under curve(AUC)of T2_wavelet_HHL_glszm_GrayLevelNonUniformity,T2_log_sigma_2_0_mm_3 D_glazm_GrayLevel-NonUniformityNormalized,ADC_wevelet_HLH_glszm_Large_AreaHighGrayLevelEmphasis for predicting prostate cancer in high and low risk groups were 0.73,0.72 and 0.71,respectively.Finally,based on the selected 10 texture features,the RF prediction model was constructed.The accuracy,sensitivity,specificity and AUC of model were 0.70,0.63,0.77 and 0.76,respectively;the positive predictive value and negative predictive value were 0.75 and 0.65,respectively.Conclusion It is demonstrated that the biparametric MRI texture feature analysis combined with machine learning model showed clinical application value in prediction of high-risk and low-risk prostate acinus cancer.