不同弥散加权成像对基于生成对抗网络的前列腺癌检测的影响

目的 探讨不同扩散敏感系数（b值）的弥散加权成像（DWI）对基于生成对抗网络（GAN）的前列腺癌（PCa）检测影响的价值。方法 回顾性收集2012年1月—2018年6月同济大学附属同济医院就诊的前列腺疾病病例446例，其中PCa有174例、前列腺增生（BPH）有272例，所有病例均采用Siemens Verio 3.0T MRI扫描并经直肠超声引导下前列腺穿刺活检或前列腺根治术后病理证实。MRI成像序列包括横断位、矢状位高分辨T2加权成像（T2WI），扩散敏感系数（b值)分别为0、500、1000s/mm2横断位弥散加权成像（DWI）及动态对比增强（DCE）扫描，通过Matlab后处理计算化合成b分别为1500、2000s/mm2的DWI图像。本研究提出一个新型神经网络模型SegDenseAN，并结合不同b值DWI图像进行检测。将不同b值DWI与ADC影像的组合作为SegDenseAN网络的输入，各组合分别为: 组合1: ADC图；组合2: ADC+DWI0+DWI500；组合3: ADC+DWI0+DWI1000；组合4: ADC+DWI0+DWI1500；组合5: ADC+DWI1000+DWI1500；组合6: ADC+DWI1000+DWI2000，分析比较不同组合对准确率的影响。结果 组合1~6的准确率分别为0.871、0.887、0.903、0.903、0.903、0.935；组合1～6的灵敏度分别为0.935、0.935、0.968、0.968、0.968、0.968；组合1~6的特异度分别为0.806、0.839、0.839、0.839、0.839、0.903；组合6的前列腺癌病灶区域识别最接近于前列腺癌标注的金标准。结论 SegDenseAN 可以实现对于病灶区域的自动分割进而有助于前列腺癌的自动检测；多b值尤其是多高b值DWI与ADC影像的不同结合对算法的检测效果有影响，多个高b值DWI图像与ADC图结合有助于提高前列腺癌的智能检测结果。

Effects of different Diffusion-Weighted Imaging in Detection of Prostate Cancer based on Generative Adversarial Networks

Objective To investigate the value of diffusion-weighted imaging(DWI) with different b values on the detection of prostate cancer(PCa) based on the generative adversarial networks(GAN). Methods A total of 446 patients with prostate disease admitted in Tongji Hospital from January 2012 to June 2018, including 174 cases of PCa and 272 cases of benign prostatic hyperplasia(BPH) were retrospectively analyzed. All patients underwent a MRI scan and were pathological confirmed by transrectal ultrasound guide biopsy or radical prostatectomy. The imaging sequence included high-resolution axial and sagittal T2 weighted imaging(T2WI), axial acquired diffusion weighted imaging(DWI) with b=0, 500, 1000s/mm2 and dynamic contrast-enhanced(DCE) scans. DWI images of b=1500, 2000s/mm2 were calculated and synthesized with Matlab post processing. A new neural network model SegDenseAN, which combined different b-value DWI images, was developed. The combination of different b-value DWI and ADC images was used as the input of SegDenseAN network. The combinations of groups 1~6 were: ADC map, ADC+DWI0+DWI500, ADC+DWI0+DWI1000, ADC+ DWI0+DWI1500, ADC+DWI1000+DWI1500 and ADC+DWI1000+DWI2000. The effects of different combinations on detection accuracy were analyzed and compared. Results The diagnostic accuracy for prostate cancer of groups 1-6 was 0.871, 0.887, 0.903, 0.903, 0.903 and 0.935, respectively; the sensitivity of groups 1-6 were 0.935, 0.935, 0.968, 0.968, 0.968, 0.968, respectively; the specificities were 0.806, 0.839, 0.839, 0.839, 0.839, and 0.903, respectively. The diagnostic value for diagnosis of prostate cancer of combination group 6 was better than that of other combinations. Conclusion SegDenseAN can realize the automatic segmentation of the lesion area and help the automatic detection of PCa. The multi-b value, especially the high b-value DWI and ADC image combination have different effects on the detection effect of the algorithm. Multiple high-b value DWI images combined with ADC images may improve the intelligent examination of prostate cancer.