基于密集连接和Inception模块的前列腺图像分割

Prostate image segmentation based on dense connections and Inception module

针对目前磁共振影像上前列腺组织区域的自动分割存在分割精度较低和过分割等问题,提出了一种基于密集连接和Inception模块的U-Net分割算法。首先采用对比度受限的自适应直方图均衡化方法对前列腺图像进行处理,增强信息的可检测性。此外,该算法将密集连接思想引入到U-Net模型中,改进原有编码器和解码器的连接方式,实现多尺度语义信息的融合和传播。同时,使用由空洞卷积驱动的Inception模块代替原有的级联卷积操作,以增加网络的宽度,增强对不同尺寸目标的特征提取与表达能力。最后,针对非组织目标存在的过分割问题,设计了一种具有分类引导功能的校正器,以减少假阳性预测。通过对NCI-ISBI 2013 Challenge公开数据集进行测试,以Dice相似系数、准确率和假阳率作为评价标准,其均值分别可达86.12%、97.96%和1.11%。实验结果表明,与其他分割算法相比,该算法具有更好的分割效果。

Aiming at the problems of low segmentation accuracy and over-segmentation in the current automatic segmentation of prostate tissue regions on magnetic resonance images, a U-Net segmentation algorithm combining dense connections and Inception modules was proposed. Firstly, the contrast-limited adaptive histogram equalization method was used to process the prostate image to enhance the detectability of the information. In addition, the algorithm introduces the idea of dense connection into the U-Net model, improves the connection method of the original encoder and decoder, and realizes the fusion and dissemination of multi-scale semantic information. Meanwhile, the Inception module driven by atrous convolution is used to replace the original concatenated convolution operation to increase the width of the network and enhance the feature extraction and expression capabilities for objects of different sizes. Finally, for the over-segmentation problem of non-organized objects, a corrector with classification-guided function is designed to reduce false positive predictions. By testing on the public dataset of NCI-ISBI 2013 Challenge, using Dice similarity coefficient, accuracy rate and false positive rate as evaluation criteria, the mean values can reach 86.12%, 97.96% and 1.11%, respectively. The experimental results show that this algorithm has better segmentation effect than other segmentation algorithms.