A Deep Siamese-Based Plantar Fasciitis Classification Method Using Shear Wave Elastography
A Deep Siamese-Based Plantar Fasciitis Classification Method Using Shear Wave Elastography
复制标题
基于深连体的剪切波弹性成像足底筋膜炎分类方法
DOI：
10.1109/access.2019.2940645
复制
发表时间：
2019-01-01
期刊：
影响因子：
3.9
通讯作者：
Wan, Mingxi
中科院分区：
文献类型：
--
作者：
Gao, Junling;Xu, Lei;Wan, Mingxi
二维剪切波弹性成像（2D-SWE）是评价足底筋膜炎（PF）的一种有效可行的方法。到目前为止，只有经验丰富的医生才能通过超声图像给出相对准确的评估，导致效率低，成本高。因此，迫切需要设计自动算法来识别这些超声图像的模式。近年来，深度学习（DL）在计算机辅助诊断（CAD）方面取得了长足的进步。然而，目前还没有研究将DL应用于PF的诊断。为了实现强大的PF分类，本文构建了一个具有多任务学习和迁移学习（DS-MLTL）的深度连体框架，该框架使用2D-SWE学习有区别的视觉特征和有效的识别功能。DS-MLTL模型包括两个VGG风格的分支和一个多任务损失，包括分类损失和暹罗损失。暹罗损失利用了不同图像的内在结构（相似性），并包含对比约束和相似约束。在我们的框架中，视觉特征和多任务损失是共同学习的，它们可以相互受益。为了有效地训练DS-MLTL模型，该模型将知识从大规模ImageNet数据集转移到PF分类任务。对于模型评估，收集足底筋膜的SWE数据集，其包含PF图案的282个图像和健康图案的60个图像。实验结果表明，DS-MLTL方法取得了良好的准确率为85.09 ± 6.67%，性能优于从B型超声和SWE中提取的人工特征。此外，DS-MLTL也获得了最好的性能相比，不同的DL模型。
Two-dimensional shear wave elastography (2D-SWE) is an effective and feasible method for plantar fasciitis (PF) evaluation. Until now, only experienced doctors have been able to give relatively accurate evaluation via ultrasound images, resulting in low efficiency and high cost. Therefore, designing automatic algorithms to recognize the pattern of these ultrasound images is urgently required. In recent years, deep learning (DL) has made considerable progress in computer- aided diagnosis (CAD). However, there have been no studies that apply DL to the diagnosis of PF. To achieve robust PF classification, this paper builds a deep Siamese framework with multitask learning and transfer learning (DS-MLTL), which learns discriminative visual features and effective recognition functions using 2D-SWE. The DS-MLTL model comprises two VGG-style branches and a multitask loss including a classification loss and a Siamese loss. The Siamese loss leverages the intrinsic structure (similarities) of different images and contains a contrastive constraint and a similar constraint. In our framework, visual features and the multitask loss are learned jointly, and they can benefit from each other. To train the DS-MLTL model effectively, the model transfers knowledge from the large-scale ImageNet dataset to the PF classification task. For model evaluation, an SWE dataset of plantar fascia, which contains 282 images of a PF pattern and 60 images of a healthy pattern, is collected. Experimental results show that the DS-MLTL method achieves favorable accuracy of 85.09 ± 6.67% and performs better than human-crafted features extracted from B-mode ultrasound and SWE. In addition, DS-MLTL also obtains the best performance compared with different DL models.