US signal and image monitoring and diagnosis using computer vision and deep learning
(English) This thesis presents a comprehensive investigation of the reliability and efficacy of Ultra-Sound (US) signals and images for remote monitoring and early diagnosis in medical applications. The main objective of this research is to explore methodologies for the extraction of features in US...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/404646 |
| Acceso en línea: | https://hdl.handle.net/2117/404646 https://dx.doi.org/10.5821/dissertation-2117-404646 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Enginyeria biomèdica Àrees temàtiques de la UPC::Informàtica |
| Sumario: | (English) This thesis presents a comprehensive investigation of the reliability and efficacy of Ultra-Sound (US) signals and images for remote monitoring and early diagnosis in medical applications. The main objective of this research is to explore methodologies for the extraction of features in US signals and images to be able to detect abnormalities or features for monitoring or as a first step for diagnosis. Both Computer Vision (CV) and Deep Learning (DL) techniques are investigated and compared. Another goal is to study the benefits of training a network in US data, which is scarce, from data from an environment with larger data sets to enhance pattern extraction in a data-scarce US environment. To address these questions, the research was divided into two parts. First, the analysis of US signals from a wearable device, thus allowing tele monitoring. Then, taking advantage of the experience in the processing of this kind of signals, the second part extends this research to the analysis of US images. In the first part, we evaluated the reliability of US signals for remote monitoring of the fetal heart rate in pregnant women. The methodology was evaluated by comparing the heart rate obtained from a pocket-sized wearable Doppler with those acquired with a cardiotocography device. The findings demonstrated the reliability of the US signals for remote monitoring. For the second part, the research focuses on the detection of some US features, using again simple US portable devices, to make extensive screenings possible, thus allowing an earlier diagnosis. We focused on US images, specifically Lung UltraSound (LUS) images, to explore their diagnostic capabilities. By developing advanced algorithms, we successfully detected early stage signs in lung pathology, highlighting the potential of US images for early diagnosis, especially in emergency and overloaded situations such as the COVID-19 pandemic or extensive screening programs. Due to the limited amount of data available, a common problem in Deep Learning systems, we investigated the transfer learning approach by training a network on heart ultrasound data and tuning it for its application in the processing of LUS data. The results revealed that the tuned network outperformed the one directly trained on the limited LUS data, highlighting the benefits of training in more data-rich LUS environments. In the comparison between CV and DL techniques for US image interpretation, we found that CV techniques exhibited higher scores for feature extraction in LUS images. However, the potential for DL to perform as well as, or even better than computer vision techniques was evident, especially in applications with access to more reliable and diverse data. Overall, this study provides valuable information on the potential and limitations of US signals and images for remote monitoring and early diagnosis, paving the way for further advancements in the field of medical imaging and diagnostics. |
|---|