Abstract
Hyperspectral imaging is one of the trending research topics these days due to its benefits of revealing distinct spectral contents and carrying rich information. The topic is expanding within the computer vision community, especially in biomedical computer vision. Human placenta tissues are validated as a good model for training novice neurosurgical surgeons since their vascular tree has a similar structure to the human brain's. Segmentation is the process of partitioning an image into multiple component regions or objects leading to change the representation of the original image. The obtain representation makes it easy to analyze the image so that many information of region of interest will be obtained for object scene annotation later. The master thesis aims to combine the hyperspectral imaging technique with several deep learning-based methods to perform semantic segmentation with the specific hyperspectral data - human placenta tissues. Placentas are donated the donors at Kuopio University Hospital, Finland and the hyperspectral images of human placenta are acquired and provided by Microsurgery center of Eastern Finland. Relied on the mechanism attention, we propose Dual Attention Band Selection (DABS) framework to reduce the dimensionality of the raw hyperspectral image before training. The segmentation training process includes two types of data pipelines, which are the image-based data pipeline and the patch-based data pipeline. While the former one keeps the whole hyperspectral image with 37 spectral bands and only reszie it to 256 x 256 x C as the input for model, the latter one cuts the hyperspectral image into 16 smaller cube with the same size as in the former one. C is denoted as the band numbers of the input image. We perform segmentation training mainly with two input sizes: 256 x 256 x 37 - the full bands and 256 x 256 x 20 - the selected bands subset from DABS. The chosen deep learning-based methods to segment artery, stroma and vein in the placenta tissue are HRNet and U-Net. Moddels' outputs are evaluated with mIoU metric in order to obtain the model with the highest mIoU score. Finally, a software is developed to visualize the segmentation result of the best model and it is going to be used for surgery training.