Right ventricle (RV) dysfunction and RV failure are critical complications in intensive care units (ICUs), particularly for patients undergoing mechanical ventilation. While transesophageal echocardiography (TEE) is the preferred imaging modality in these settings, current assessment methods rely on manual, offline analysis, preventing continuous monitoring of RV function, and current automated methods based on deep learning (DL) focus only on transthoracic echocardiography (TTE). This study introduces AutoRV, a fully automated deep learning pipeline for real-time RV monitoring using 2D TEE mid-esophageal 4-chamber view images. The system was developed to automatically compute clinically relevant indices of RV function based on the automatic detection of three anatomical landmarks: the tricuspid annulus (TA) hinge points, septal and free wall, and the RV apex. For the detection task, several convolutional neural networks, including ResNet variants, ResNeXt50, SwinUNETR, and U-Net, were trained, tested and validated on a dataset of 8,424 images obtained from 52 mechanically ventilated patients. The U-Net architecture showed superior performance, with a mean landmark localization error of 4.16 ± 3.36 mm on the test set. Specifically, the detection of the free-wall TA point was highly accurate, with an error of 2.93 ± 1.80 mm. The pipeline proved highly efficient, operating at an inference speed of 248 frames per second (fps), making it suitable for real-time monitoring. AutoRV automatically computes clinical indices with acceptable consistency versus manual measurements: the system achieved bias and standard deviation (std) of 1.09 ± 2.25 mm for free wall tricuspid annular plane systolic excursion (TAPSEfw) and 0.19 ± 6.64% for RV fractional area change (RVFAC). Other indices showed higher variability, particularly those involving the RV apex and the septal hinge point, as those predictions had worse results. AutoRV represents the first end-to-end system capable of automated RV assessment in TEE. Its high inference speed and ability to derive standard clinical parameters demonstrate its potential as a tool for continuous cardiac monitoring in critical care environments.
La disfunzione e l’insufficienza del ventricolo destro (RV) sono complicanze rilevanti nelle terapie intensive, soprattutto nei pazienti sottoposti a ventilazione meccanica. L’ecocardiografia transesofagea (TEE) è la tecnica di imaging di riferimento, ma le attuali valutazioni si basano su analisi manuali offline, limitando il monitoraggio continuo della funzione del RV. I metodi automatizzati basati su deep learning sono stati sviluppati quasi esclusivamente per l’ecocardiografia transtoracica. Questo lavoro presenta AutoRV, sistema completamente automatizzato basato su deep learning, progettato per il monitoraggio in tempo reale del RV tramite immagini TEE 2D in sezione medio-esofagea. Il modello è stato addestrato e validato su 8,424 frame provenienti da 52 pazienti ventilati. Sono state confrontate diverse architetture di reti neurali convoluzionali, tra cui varianti di ResNet, ResNeXt50, SwinUNETR e U-Net, per individuare tre landmark clinicamente rilevanti: i punti di inserzione settale e della parete libera dell’annulus tricuspidale (TA) e l’apice del RV. Tra le architetture valutate, U-Net ha ottenuto le migliori prestazioni, con errore medio di localizzazione pari a 4.16 ± 3.36 mm sul set di test. Il punto di inserzione del TA sulla parete libera è stato identificato con elevata precisione (2.93 ± 1.80 mm). La pipeline è computazionalmente efficiente, con velocità di inferenza di 248 frame/s, adatta al monitoraggio continuo in tempo reale. AutoRV calcola automaticamente gli indici clinici principali, con bias contenuto rispetto alle misurazioni manuali: 1.09 ± 2.25 mm per TAPSEfw e 0.19 ± 6.64% per RVFAC. Altri parametri mostrano maggiore variabilità, soprattutto quelli dipendenti dall’apice del RV e dal punto di inserzione settale, coerentemente con le prestazioni inferiori su questi landmark. AutoRV è il primo sistema end-to-end in grado di fornire valutazione automatizzata del RV da immagini TEE. La sua elevata velocità di inferenza e la capacità di estrarre oggettivamente parametri clinici standard ne evidenziano il potenziale per il monitoraggio cardiaco continuo in terapia intensiva.
AutoRV: a real-time, fully automated pipeline for RV evaluation based on transesophageal echocardiography in intensive care units
Missana, Matteo
2025/2026
Abstract
Right ventricle (RV) dysfunction and RV failure are critical complications in intensive care units (ICUs), particularly for patients undergoing mechanical ventilation. While transesophageal echocardiography (TEE) is the preferred imaging modality in these settings, current assessment methods rely on manual, offline analysis, preventing continuous monitoring of RV function, and current automated methods based on deep learning (DL) focus only on transthoracic echocardiography (TTE). This study introduces AutoRV, a fully automated deep learning pipeline for real-time RV monitoring using 2D TEE mid-esophageal 4-chamber view images. The system was developed to automatically compute clinically relevant indices of RV function based on the automatic detection of three anatomical landmarks: the tricuspid annulus (TA) hinge points, septal and free wall, and the RV apex. For the detection task, several convolutional neural networks, including ResNet variants, ResNeXt50, SwinUNETR, and U-Net, were trained, tested and validated on a dataset of 8,424 images obtained from 52 mechanically ventilated patients. The U-Net architecture showed superior performance, with a mean landmark localization error of 4.16 ± 3.36 mm on the test set. Specifically, the detection of the free-wall TA point was highly accurate, with an error of 2.93 ± 1.80 mm. The pipeline proved highly efficient, operating at an inference speed of 248 frames per second (fps), making it suitable for real-time monitoring. AutoRV automatically computes clinical indices with acceptable consistency versus manual measurements: the system achieved bias and standard deviation (std) of 1.09 ± 2.25 mm for free wall tricuspid annular plane systolic excursion (TAPSEfw) and 0.19 ± 6.64% for RV fractional area change (RVFAC). Other indices showed higher variability, particularly those involving the RV apex and the septal hinge point, as those predictions had worse results. AutoRV represents the first end-to-end system capable of automated RV assessment in TEE. Its high inference speed and ability to derive standard clinical parameters demonstrate its potential as a tool for continuous cardiac monitoring in critical care environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Missana_Tesi.pdf
accessibile in internet per tutti
Descrizione: Article format Thesis
Dimensione
5.59 MB
Formato
Adobe PDF
|
5.59 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Missana_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
1.18 MB
Formato
Adobe PDF
|
1.18 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/252505