Machine Learning in 3D Shapes: Intelligent Analysis and Prediction
Open AccessThe advancements in the optical 3D surface scan and reconstruction techniques have brought new opportunities for improving assessment by intelligently analyzing shapes. The 3D shapes are widely used to predict parameters such as the quality of dance, obesity level, pregnancy, etc. The functionality and dimensionality of predicted parameters significantly impact the models. Currently, many studies apply 3D shapes to predict parameters. However, no specific research focuses on the difference in the predicted parameters. Therefore, this dissertation proposes novel machine learning models (statistical and deep learning) to predict various responses (i.e., the categorized, continuous, and high dimensional) with 3D shapes.The statistical model is easy to implement and always the first choice of predicting the parameters. We present a statistical model which includes two steps: one uses functional data analysis to select an optimal region of interest (ROI), and the second is to model the ROI. We present a highly innovative strategy in selecting the ROI of shape to reduce the overfitting risk. We evaluate the overfitting risk of each candidate sub-interval by calculating the necessary sample size to achieve a pre-specified statistical power. Combining with a model accuracy measure, we rank these sub-intervals and select the optimal sub-interval/ ROI. After that, we propose a shape-based prediction model and apply it to visceral adipose tissue assessment. We train the baseline prediction model for males and females, respectively. And then, we integrate a gender likelihood model and a shape prior model into the baseline model to reduce the impacts of different gender and shape types. Deep CNN is good at solving classification problems. Thus we present a deep convolutional neural network (CNN) model to evaluate categorized parameters. We use the 2D depth maps of 3D shapes to construct a classification network. The proposed network adopts the dilated residual network blocks to extract the refined features of the inputs by expanding the receptive field. Furthermore, we create a hybrid of the center loss and cross-entropy loss to compact intra-class variations and separate inter-class differences. We test our model on hepatic steatosis assessment, and our experimental results show that the proposed network outperforms competitive methods. We also present a deep regression network with an attention module to assess the continuous parameters. The proposed model comprises a flexible baseline network and a lightweight multi-channel attention module. The attention module is trained to generate discriminative and diverse features, significantly improving performance. We perform extensive tests on the liver fat percentage estimation to validate the method. The results verify that our proposed method yields state-of-the-art performance.The generative model works well in producing high dimensional data. We present a multi-task conditional generative adversarial network to predict the high dimensional parameters. We also introduce an interpreted patch discriminator to optimize the regional accuracy. We apply our model to the fat distribution prediction. Our proposed approach outperforms competitive methods with whole-body fat percentage and regional fat predictions.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.