| Abstract | Near infrared fluorescence optical imaging (NIR-FOI) visualizes the vascular perfusion of the investigated anatomical structure. Even though there has been a lot of medical research in the field to detect joint inflammation utilising NIR-FOI, an objective machine learning based evaluation method of the image data has not been developed, yet. The measured NIR-FOI data consists of two spatial dimensions (image pixel) and one temporal dimension. To assess the distribution process an understanding of the hands’ locations is essential. However, random motion changes the positioning, which requires re-segmentation. The goal of this work is to identify the time points (frames) and severity of motion in the previously measured image stack. Due to properties of the NIRFOI, each data set is split into two phases: Before and after full illumination of the hands. For each phase, an independent model is trained to evaluate the severity and time point of possible motion. The model for the first phase achieves a precision of 20.78% and a recall of 69.57 %, while the model for the second phase reaches a precision of 67.71% and a recall of 98.49% to detect non-negligible motion. Despite low precision, both models can be considered a success, contemplating the high heterogeneity, self-illumination and real-life consequences of a low precision value, which only affects computation time. Our general goal is to achieve a robust and early detection of psoriatic arthritis, to increase quality of life while decreasing treatment costs. The presented work plays a key role in this research, especially increasing robustness of the final evaluation pipeline. |
|---|
