基于卷积神经网络经直肠超声模型预测前列腺癌Gleason分级

Transrectal ultrasonography model based on convolutional neural network for predicting Gleason score of prostate cancer

目的观察基于卷积神经网络(CNN)的经直肠超声(TRUS)模型预测前列腺癌Gleason分级(GS)的价值。方法前瞻性收集101例拟接受TRUS引导下前列腺穿刺活检的前列腺癌患者;采集568幅前列腺癌超声图像,根据病理结果将其分为低危(GS≤6,n=90)、中危(GS=7,n=185)及高危(GS≥8,n=293)。建立前列腺癌TRUS数据集,基于CNN构建TRUS预测前列腺癌GS模型;以穿刺活检病理结果为金标准,评估模型与超声医师的诊断效能。结果基于CNN的TRUS模型预测前列腺癌GS≤6的精确率高于超声医师(P<0.05),而二者召回率(Recall)和F1-score差异均无统计学意义(P均>0.05);基于CNN的TRUS模型预测GS=7和GS≥8的精确率、Recall及F1-score均高于超声医师(P均<0.05)。基于CNN的TRUS模型预测前列腺癌GS的总体准确率(76.75%)高于超声医师(51.75%,χ^(2)=31.021,P<0.001),其预测前列腺癌GS的曲线下面积(AUC)为0.72、特异度为47.22%、敏感度为96.88%,超声医师分别为0.67、52.78%及80.21%,二者AUC差异无统计学意义(Z=0.859,P=0.390)。结论基于CNN的TRUS模型有助于预测前列腺癌、尤其需要积极治疗的中-高危前列腺癌的GS。

Objective To observe the value of transrectal ultrasonography(TRUS)model based on convolutional neural network(CNN)for predicting Gleason score(GS)of prostate cancer.Methods A total of 101 patients with prostate cancer who would undergo TRUS-guided prostate biopsy were prospectively enrolled.Then 568 ultrasound images of prostate cancer were collected and divided into low-risk(GS≤6,n=90),medium risk(GS=7,n=185)or high-risk(GS≥8,n=293)according to pathological results.TRUS dataset of prostate cancer was established,and TRUS prostate cancer GS model was constructed based on CNN.Taken pathological results of puncture biopsy as the gold standards,the diagnostic efficacy of this model and ultrasound physicians were analyzed.Results The precision of TRUS model based on CNN for predicting GS≤6 prostate cancer was higher than that of ultrasound physicians(P<0.05),while there was no significant difference of Recall nor F1-score between model and physicians(both P>0.05).The precision,Recall and F1-score of the model for predicting GS=7 and GS≥8 were all higher than those of ultrasound physicians(all P<0.05).The overall accuracy of TRUS model based on CNN for predicting GS of prostate cancer was higher than that of ultrasound physicians(76.75%vs.51.75%,χ^(2)=31.021,P<0.001).The area under the curve(AUC)of TRUS model based on CNN for predicting GS of prostate cancer was 0.72,with specificity of 47.22%and sensitivity of 96.88%,while of ultrasound physicians was 0.67,52.78%and 80.21%,respectively,and there was no significant difference of AUC between model and physicians(Z=0.859,P=0.390).Conclusion TRUS model based on CNN was helpful for predicting GS of prostate cancer,especially for middle and high risk prostate cancer needed to be actively treated.