Accurate segmentation of 2D echocardiographic images has been a recurring problem for over 30 years and remains the subject of intense research by the entire scientific community today. The precise delineation of cardiac structures, in fact, represents a common and crucial clinical task for establishing a correct diagnosis. The very nature of echocardiographic images makes it difficult to accurately locate cardiac regions. Furthermore, the lack of large-scale publicly available 2D echocardiographic datasets and the absence of multi-expert annotations on large datasets to evaluate the minimum error margin beyond which segmentation methods would be considered as accurate as human experts have limited the possibility of experimentation. From these observations, the CAMUS project (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation) was born; it is the largest publicly available and fully annotated dataset for echocardiographic evaluation, composed of two- and four-chamber acquisitions from 500 patients, with measurements of left ventricular volumes at end-diastole and end-systole. The objective of this study is to use this dataset to evaluate the capabilities of methods based on deep convolutional neural networks in analyzing this type of images. To this end, we implemented and tested two specific architectures, namely the Fully Convolutional Network (FCN) and the U-Net, using the open-source library PyTorch. The performance of the models was evaluated using specific metrics such as Intersection Over Union (IoU) and the Dice coefficient, which provide measures of accuracy and similarity between the model predictions and the real labels. The results obtained at the end of all the experiments conducted demonstrated that encoder-decoder-based architectures are particularly suitable for these purposes, especially the optimized U-Net, which shows excellent segmentation values, faithfully reproducing expert analysis for left ventricular volumes at end-diastole and end-systole.
La segmentazione accurata delle immagini ecocardiografiche 2D è un problema ricorrente da oltre 30 anni, ancora oggi oggetto di intensa ricerca da parte di tutta la comunità scientifica. La delineazione puntuale delle strutture cardiache, infatti, rappresenta un compito clinico comune e cruciale per stabilire una diagnosi corretta. La natura stessa delle immagini ecocardiografiche rende difficile localizzare accuratamente le regioni cardiache. Inoltre, la mancanza di set di dati ecocardiografici 2D su larga scala disponibili al pubblico e la mancanza di annotazioni multi-esperto su grandi set di dati per valutare il margine di errore minimo oltre il quale i metodi di segmentazione sarebbero considerati accurati quanto gli esperti umani hanno limitato la possibilità di sperimentazione. Da queste osservazioni nasce il progetto CAMUS (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation), il più grande dataset pubblico e completamente annotato per la valutazione ecocardiografica, composto da acquisizioni a due e quattro camere di 500 pazienti, con misurazioni dei volumi ventricolari sinistri a fine diastole e a fine sistole. L’obiettivo di questo studio è quello di utilizzare il dataset CAMUS per valutare le capacità di metodi basati su reti neurali convoluzionali profonde nell’analizzare questo tipo di immagini. A tal fine, abbiamo implementato e testato due architetture specifiche, ovvero la Fully Convolutional Network (FCN) e la U-Net, utilizzando la libreria open source PyTorch. Le performance dei modelli sono state valutate tramite metriche specifiche come l'Intersection Over Union (IoU) e il Dice coefficient, che fornisce una misura dell’accuratezza e della similarità tra le previsioni del modello e le etichette reali. I risultati ottenuti alla fine di tutti gli esperimenti eseguiti hanno dimostrato come le architetture basate su encoder-decoder siano particolarmente adatte a questi scopi, in particolare la rete U-Net ottimizzata mostra ottimi valori di segmentazione, riproducendo fedelmente l’analisi degli esperti per i volumi ventricolari sinistri a fine diastole e fine sistole.
Modelli di deep learning per la segmentazione automatica delle immagini ecografiche
NATO, GIUSEPPE
2023/2024
Abstract
Accurate segmentation of 2D echocardiographic images has been a recurring problem for over 30 years and remains the subject of intense research by the entire scientific community today. The precise delineation of cardiac structures, in fact, represents a common and crucial clinical task for establishing a correct diagnosis. The very nature of echocardiographic images makes it difficult to accurately locate cardiac regions. Furthermore, the lack of large-scale publicly available 2D echocardiographic datasets and the absence of multi-expert annotations on large datasets to evaluate the minimum error margin beyond which segmentation methods would be considered as accurate as human experts have limited the possibility of experimentation. From these observations, the CAMUS project (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation) was born; it is the largest publicly available and fully annotated dataset for echocardiographic evaluation, composed of two- and four-chamber acquisitions from 500 patients, with measurements of left ventricular volumes at end-diastole and end-systole. The objective of this study is to use this dataset to evaluate the capabilities of methods based on deep convolutional neural networks in analyzing this type of images. To this end, we implemented and tested two specific architectures, namely the Fully Convolutional Network (FCN) and the U-Net, using the open-source library PyTorch. The performance of the models was evaluated using specific metrics such as Intersection Over Union (IoU) and the Dice coefficient, which provide measures of accuracy and similarity between the model predictions and the real labels. The results obtained at the end of all the experiments conducted demonstrated that encoder-decoder-based architectures are particularly suitable for these purposes, especially the optimized U-Net, which shows excellent segmentation values, faithfully reproducing expert analysis for left ventricular volumes at end-diastole and end-systole.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_Nato.pdf
non disponibili
Dimensione
12.06 MB
Formato
Adobe PDF
|
12.06 MB | Adobe PDF | Richiedi una copia |
È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/33329