Ciencias Exactas y Ciencias de la Salud

The application area of this work is mainly on the case study of the identification of kidney stones. Currently the standard identification process involves extraction of the stone, surface and sectional visual inspection of it and a Fourier-Transform Infrared spectroscopy (FTIR) analysis. This process is known as Morpho-Constitutional Analysis (MCA) and currently it can take up to a couple of weeks or even a month. Since identification is essential to prescribe a treatment and in some cases formation of new kidney stones can be in only a couple of weeks there is the need to develop a faster and reliable method of identification of kidney stones. We want to get recommended classifications of the endoscopic images of kidney stones that are easy to analyze. Those classifications should be produced along with pertinent information about the causes of such classifications. This information should allow specialists to confirm if the kidney stones were or were not correctly classified. Then, a Computer Assisted Diagnosis (CADx) tool for the classification of kidney stones is the goal pursued. Recently, the Deep Learning (DL) field has shown good results in many different areas, but at the expense of relying on models with millions or billions of parameters. As a result, it impedes human interpretation of the behavior of these models. In the literature, this limitation is frequently mentioned as the Black-Box nature of DL models. Due to the high importance of mitigating this aspect in the adoption of DL models in this work self-explainable methods are explored. For this thesis, convolutional neural networks are leveraged as feature extractors from images on which prototypical parts learned by a DL model will identify their level of similarity to determine the presence or absence of the characteristic parts that compose certain image classes. This approach is said to be considered self-explainable since the model will identify the relevant parts of an image that are considered to be due to the presence of a certain type of entity, which is been classified, and based on these detections, generate visualizations that the model itself deems the most similar previously to the final classification of the input image. This behavior, hence, allows us to visualize and corroborate if the identified parts indeed correspond to parts relevant and indicative of the type of image the DL model concludes is the class of the input image. Additionally, the levels of similarity and their example cases are evaluated to determine the main visual characteristics behind those models' activations and their possible causal relationship to the model's final output. You can find the base code for the development of the different experiments explored in this dissertation at: https://github.com/DanielF29/PPs_ICNN_Loss

Ciencias Exactas y Ciencias de la Salud

Browse

Filters

Settings

Sort By

Results per page

Search Results