Corruption identification and classification in standard plane prenatal ultrasounds
Tipus de documentProjecte Final de Màster Oficial
Data2020-09-09
Condicions d'accésAccés obert
Llevat que s'hi indiqui el contrari, els
continguts d'aquesta obra estan subjectes a la llicència de Creative Commons
:
Reconeixement-NoComercial-CompartirIgual 3.0 Espanya
Abstract
Artificial Intelligence (AI) has proven in the recent years to be a powerful tool to automate simple and complex tasks. One of the main fields where it has been applied is medical science where especially computer vision algorithms are gaining popularity. This project aims to be useful both for medical staff and future projects on the field, easing the process of data cleaning. Nowadays medical personnel have to invest time discarding corrupted images to clean the data from the ultrasounds. It is also an impediment for other machine learning projects as corrupted images can negatively affect the results. Our goal is to develop a Deep Learning (DL) algorithm not only able to identify the corrupted images but also to classify them by type of distortion. With this we will be able to clean the dataset and bring an opportunity to future projects to apply artifact cleaning. The proposed strategy has three main parts. First, we have designed a code to create semi random artifacts on correct images to generate a dataset big enough for our experiments. Secondly, we have studied the behavior under corrupted images of a previous deep learning model, designed by Vicent at [6]. With this we determined the boundary between corruption and correct images. Finally, we successfully created and compared the result of two models based on a Convolutional Neural Networks (CNNs) to identify and classify the corrupted images.
TitulacióMÀSTER UNIVERSITARI EN ENGINYERIA INDUSTRIAL (Pla 2014)
Col·leccions
Fitxers | Descripció | Mida | Format | Visualitza |
---|---|---|---|---|
memoria-tfm-lluis-cierco-corominas.pdf | 2,132Mb | Visualitza/Obre | ||
annex-tfm-lluis-cierco-corominas.pdf | 447,8Kb | Visualitza/Obre |