Machine learning approaches for slice to volume biomedical data integration

Abstract

Imaging plays an essential role in modern biomedical sciences and lays the foundation for current research and clinical diagnosis. During the last decade, slice-to-volume registration, a particular type of image registration problem, has received great attention from the medical imaging community due to the emergence of several medical applications of slice-to-volume mapping (2D to 3D image mapping) using biomedical data such as biomedical atlas. The task of integrating new data into a biomedical atlas is a typical 2D to 3D image registration problem. Images created in experiments are mostly 2D images, while modern biomedical atlases are mostly 3D models. To transfer the data related to the 2D image (e.g., gene expression data) to the 3D Atlas, it is necessary to determine the position of the new image in the 3D model. This is typically done by experts who review the 2D sections and manually position 2D data into 3D with some tools. Manual positioning 2D data into 3D is financially expensive, time consuming, and require extensive work by experts. However, finding experts who have domain knowledge is also another crucial challenge. To resolve this problem, this thesis automate the process of positioning the 2D image into the 3D model. This study contributes by creating two datasets that convert the 3D Atlas into a series of 2D slices. Then, we utilize a Convolutional Neural Network (CNN) for registering purposes. The proposed CNN model is trained to determine the distance and pitch values used to describe the position of the 2D slice in the atlas coordinate system, and the proposed model obtained 94% accuracy. Furthermore, we tested different variants of CNN architectures and different transfer learning techniques to build an optimal image base model for image analysis. We employ all the data modalities available in the biomedical Atlases, such as the images and the textual anatomical data. To test the performance in real-life situation, the performance of the proposed model is evaluated on the unseen dataset. The results show that the proposed model outperforms the image-only data and obtain 97% accuracy. A different data set (contained cropped images) is used to test the performance of the proposed technique for image matching, and the algorithm achieved 94% accuracy. The study has shown that different data modalities available within the atlases can train the machine learning to overcome many of the issues related to the use of image-processing based or ontology-based techniques.