Monitoring and Classification of Tissue Perfusion Through the Combined Use of Thermal, RGB, and Laser Speckle Contrast Imaging

Abstract

In cooperation with the UNLV School of Medicine, this thesis is part of a bigger project aimed at developing a post-op surgery device for the constant unattended monitoring of a patient’s tissue. This device is intended to monitor areas of surgical sites and detect any complications without the need for continuous supervision by medical professionals. One such complication includes inadequate tissue perfusion, or a lack of blood flow reaching a patient’s tissue. The role of this thesis is to fit a convolutional neural network model to the available data gathered from voluntary participants and attempt to classify the data as a different stage of tissue perfusion in the hand area. The device is equipped with a camera setup using three sensing modalities, RGB, thermal infrared, and laser speckle contrast imaging, revealing information that a single conventional camera could not. Once the data is prepared for model training, various techniques will be explored due to the multiple sensing modalities involved in data collection. The techniques range from the training of a single image modality to the topic of data fusion, a key component of the thesis. In the early fusion implementation, the data to be fused is the image channels coming from the 3 modalities, creating a 4 to 5 channel image. The late fusion implementation will train each modality separately, and fuse feature vectors at the end of the model. This thesis will present the implementation and results of both techniques.