基于扩张卷积与注意力的甲状腺超声分割方法-《计算机技术与发展》

文章信息/Info

Title:: Dilated Convolution and Attention-based Ultrasound Segmentation of Thyroid Nodules

Author(s):: XIE Zi-wei; LU Da-ying* ; LI Zhi-qi; KONG Chen-xi; WU Xi; ZHANG Jun; School of Cyberspace Security,Qufu Normal University,Qufu 273100,China

Keywords:: thyroid nodules; ultrasound images; dilated convolution; channel attention; deep learning

摘要:: 甲状腺结节是常见的内分泌疾病,从超声图像中准确分割出结节是一项重要工作。为了有效地解决原始超声图像噪声大、对比度低、结节与周围组织互相粘连的问题,呈现出结节清晰的轮廓形态,提出一种基于深度学习的甲状腺超声结节分割方法。利用扩张卷积模块增加分割模型的感受野范围并且保持特征图的尺寸不变,精准提取更广阔的上下文信息;构建高效通道注意力机制模块,动态地调整通道特征权重,突显出超声图像中的重要关键信息;并且设计混合双损失函数来保障模型的性能和分割的准确性。将此方法应用到甲状腺数据集上进行消融实验验证各模块有效性,同时与已有的方法在多个评价指标上进行比较,结果表明,该方法的精确度和 F1-Score 可分别达到 0. 971 2 和 0. 971 5,与其他经典方法相比可以更精确地分割甲状腺结节。

Abstract:: Thyroid nodules are common endocrine diseases,and accurate segmentation of nodules from ultrasound images is an importanttask. In order to effectively solve the problems of high noise,low contrast and adhesion of nodules to surrounding tissues in the originalultrasound image and present a clear contour morphology of nodules, a deep learning based thyroid nodule segmentation method isproposed. The dilated convolution algorithm is used to increase the perceptual field of the ultrasound image to keep the size of the featuremap constant and accurately extract the broad contextual information. An efficient channel attention mechanism module is constructed todynamically adjust the channel feature weights to highlight the important key information in the ultrasound image,and a hybrid dual lossfunction is designed to guarantee the accuracy of the method. This method is applied to the thyroid dataset for ablation experiments toverify the validity of each module,and compared with existing methods on several evaluation metrics. It is showed that the precision andF1-Score of the proposed method can reach 0. 971 2 and 0. 971 5，respectively,which can segment thyroid nodules more accuratelycompared with other classical methods.