medical processing

The classification of Doppler ultrasound imagesis very important for conception prediction. However it is achallenging problem that suffers from a variable length of thoseimages with a dimension gap between them. In this study, wepropose a latent representation weight learning method (LRWL)for conception prediction with Doppler ultrasound images. Unlikemost existing related methods, LRWL can process a variablelength of multiple images, particularly with an irregular multi-image issue. LRWL can extract the relation between the imagesand then learn the latent representative weight of each image.Comparatively, we also propose another method—spatiotemporalinteraction measurement (SIM), to validate the experimentalassumption that LRWL can describe the role of each image moreaccurately. Then we integrate the images with the weights andthe diagnostic indices, as the input data to a deep learning (DL)model to predict the successful conception. Finally, we conductthe comprehensive experiments with the classification tasks onthe real irregular reproduction datasets and two other syntheticregular datasets. Experimental results show that the proposedLRWL outperforms existing relative methods and is suitable forthe irregular multi-image datasets. It can be efficiently optimizedunder the limited-memory Broyden-Fletcher-Goldfarb-Shannobound constraint algorithm (L-BFGS-B), and the alternatingdirection minimization (ADM) framework. Thus, the proposedmethod can achieve the good performance with high accuracyand good convergence.

多普勒超声图像（Doppler ultrasound images）的分类对于受孕预测至关重要，但这类图像存在长度不一、维度存在差异的问题，是一项极具挑战性的任务。本研究提出了一种面向多普勒超声图像受孕预测的隐式表征权重学习方法（LRWL）。与多数现有相关方法不同，LRWL能够处理多幅长度不一的图像，尤其适用于不规则多图像场景。该方法可提取图像间的关联关系，并学习每幅图像的隐式表征权重。作为对照，本文还提出了时空交互度量方法（SIM），以验证LRWL能更准确地描述每幅图像的作用这一实验假设。随后将带权重的图像与诊断指标相结合，作为深度学习（DL）模型的输入数据，用于预测受孕成功与否。最后，本文在真实的不规则生殖数据集以及另外两个合成规则数据集上开展了分类任务的综合实验。实验结果表明，所提LRWL方法优于现有相关方法，适用于不规则多图像数据集。该方法可在有限内存布洛登-弗莱彻-戈德法布-尚诺边界约束算法（L-BFGS-B）与交替方向极小化（ADM）框架下高效优化，因此能够实现高精度且收敛性良好的优异性能。