Adopting Signal Processing Technique for Osteoporosis Detection Based on CT Scan Image

Abstract

Machine learning (ML) and artificial intelligence (AI) are widely applied in many disciplines including medicine. Pattern recognition or automatization has been successfully implemented in various field studies. Similarly, multiple efforts have been made in medicine to implement AI/ML technology to solve medical problems, for example, for automating osteoporosis detection. In general, the success of AI/ML technology is highly dependent on the amount of available data, especially during the training stage. Feature generation is a common technique that allows the manipulation of available data for the training stages. This paper aims to study the feasibility of adopting signal-processing techniques for feature generation in medical image processing. Signal attributes from signal processing workflow were adopted and applied to image processing of CT and DEXA scanning data to differentiate between normal and osteoporotic bone. Five attributes, namely amplitude, frequency, instantaneous phase, roughness, and first derivative or contrast attributes, have been tested. An attribute index number is formulated to indicate the attribute’s strength at the selected region of interest (ROI). A case study applying these attributes to the CNN model is presented. More than five hundred CT scan images of normal and osteoporosis bone were used during the training stage to test classification performance with and without developed attributes as an input. From the ten selected CT scan images used to test the CNN model, 90% were well predicted in the scenario only utilizing the grayscale as input. However, when including the developed attributes, the CNN can predict all the images well (100% were well predicted). In conclusion, the technique adopted from the signal-processing technique has the potential to enhance feature generation in image processing, whereby the results can be used for the early application of AI/ML in osteoporosis identification. Further research testing this proposed method in different image modalities needs to be conducted to verify the robustness of the proposed method.