Preliminary Domain Adaptive Algorithm for Sarcopenia Identification

Sarcopenia is a progressive skeletal muscle disorder, characterized by the loss of muscle mass, strength and function and it is typically found in elderly people. It has been gaining increasing interest in the past decade, as the first consensus was reached in 2010, when the European Working Group on Sarcopenia in Older People (EWGSOP) published the clinical definition of this illness, and was officially recognized as a musculoskeletal disease in 2017. Yet, there is still ongoing debate on how to diagnose the illness: no common agreement on the variable to measure among muscle mass, strength or quality; different cut-off points for detecting the illness; limitations and downsides in the current assessment tools. In fact, various tools exist to diagnose sarcopenia, such as CT-scan, MRI and Ultrasound. However, all these devices have some issues in terms of expense of the equipment, required expertise, portability, measurement time and patient discomfort. These are the reasons that inspired the Eurostars project MAS – Muscle Analyzer System (Ref. 2020-03595). Its goal is the development of an inexpensive, portable, quick and easy-to-use device, that, using microwave-based sensors, is able to effectively assess muscle quality. Two main transmission sensors, namely SRR and BF, were already developed in this context, however there is still ongoing debate on which one is best suited, since both have shown correlation with the muscle quality. Moreover, there is still no algorithm that is able to extract the muscle quality from the sensors output, given also the various sources of uncertainty during measurements. This work aims at developing an initial algorithm to do so, by designing a DL architecture that understands the correlation between the complex sensor measures and the muscle quality, which then uses to output a muscle quality index. This thesis still represents a preliminary stage of the whole MAS system and it is used as prototype to better understand the current sensitivity of the MAS device, its correlation with the muscle quality and to evaluate the performance difference among sensors. Given the lack of data and the unfeasibility to obtain large amounts, DL based generative data augmentation was firstly applied. In particular, a Variational Autoencoder was developed, in order to generate synthetic data and enrich the dataset with artificial patients, improving so the robustness of the whole model. A classic neural network structure was developed, which regresses the muscle quality index using real and augmented data. The results showed a lot of variability among patients and overall not good performances on one sensor. Therefore, DL-based domain adaptation was implemented, which extracts only the muscle-related information from the input data and removes the patient-dependent features. In this way, the model was able to better generalize on unseen patients the relationship learnt on training patients. Various approaches of this technique were tested and compared to evaluate the best suited in this scenario. The whole model was separately trained on both sensors, in order to evaluate the error of each one and provide results to help for the final decision on the sensor choice. The model was tested on every patient, displaying lower variability in both cases. It was found that both sensors have significant correlation with the measured muscle quality, with the SRR sensor seemingly more accurate, influenced also by a better dataset structure and a more accurate target muscle index.