基于Transformer和生成对抗网络的临床靶区分割方法

Clinical target volume segmentation method based on Transformer and generative adversarial network

针对前列腺临床靶区的对比度不均、边缘模糊、形态大小不一等问题,提出了一种基于Transformer和生成对抗网络的前列腺临床靶区分割方法(TG-UNet)。首先借助多头注意力机制有效地提取全局及局部信息,将分割结果输入到生成对抗网络,然后判别网络判断输入是分割预测结果还是真实标签,最后分割网络利用判别结果优化调整自身参数以获得更贴近真实标签的分割结果。通过多组对比实验可知,与U-Net相比,该网络在Dice系数、平均交并比(mIOU)、平均像素准确度(mPA)等指标上分别提升了9.98个百分点、12.41个百分点、0.68个百分点。实验结果表明,所提方法能够提取更多前列腺临床靶区细节,且具有较强的泛化能力。

Aiming at the problems of uneven contrast,blurred edges and varying morphological sizes of the prostate clinical target volume,A clinical target volume segmentation method based on Transformer and Generative adversarial Network(TG-UNet)was proposed.Firstly,multi-headed attention mechanism was used to effectively extract global and local information.Then,the segmentation result was input to the generative adversarial network,and the discriminative network judged whether the input was a segmentation prediction result or a true label.Finally,the segmentation network was used to optimally adjust its parameters according to the discriminant results,for obtaining a segmentation result closer to the true label.Through multiple comparison experiments,it can be seen that compared with U-Net,the proposed network improves the mean Dice coefficient,the mean Intersection-Over-Union(mIOU),and the mean Pixel Accuracy(mPA)by 9.98 percentage points,12.41 percentage points,and 0.68 percentage points,respectively,indicating that the proposed method can extract more details of the clinical target volume of the prostate and has strong generalization ability.