In the context of fetal neurosonography, the detection of standard planes plays a fundamental role, as the analysis of these specific views enables the monitoring of fetal anatomical structures throughout gestational age and provides key measurements for the assessment of potential abnormalities. Artificial intelligence, and in particular convolutional neural networks (CNNs), is extensively applied in the acquisition of standard planes. In the literature, this task has generally been addressed as a multiclass classification problem. In this work, the problem is approached from a localization perspective, explicitly focusing on the identification of anatomical landmarks within ultrasound images. To this end, a multitask multilabel-detection framework is proposed, in which the detection branch serves as an auxiliary component to the multilabel classification task, providing additional spatial awareness during training. Building upon a baseline multilabel classification model, the proposed framework is evaluated by comparing its classification metrics with those of the multitask approach. The architecture is based on a ResNet101 backbone connected to a classification head and a detection head (adapted from yolo). The loss function combines classification and detection objectives, allowing the backbone weights to be updated jointly during backpropagation across both tasks. Experimental results demonstrate strong performance and tangible improvements in classification metrics with the multitask framework compared to the multilabel baseline. Notably, the F1-score consistently exceeds 0.90 across all classes, confirming the effectiveness and robustness of the proposed method.
Nell'ambito della neurosonografia fetale, il rilevamento dei piani standard ricopre un ruolo fondamentale, perché dall'analisi di queste viste particolari risulta possibile monitorare le strutture anatomiche del feto durante l'età gestazionale ed è possibile registrare misure importanti per valutare la presenza di eventuali anomalie. L'intelligenza artificiale, attraverso l'uso di CNN è ampiamente usata per l'acquisizione di piani standard. In letteratura questo compito è stato analizzato come un compito di classificazione multiclasse. In questo progetto si affronta il problema da una prospettiva di localizzazione mirando esplicitamente all’identificazione dei landmark all’interno delle immagini. Si propone dunque un approccio multi-task di classificazione multilabel-detection in cui il rilevamento funge da ausiliario al multilabel conferendogli maggiore consapevolezza spaziale in fase di addestramento. Sulla base di una classificazione multilabel classica, si confrontano le metriche con il framework multitask per valutarne l'efficacia. Il backbone usato per la definizione dell'architettura del multitask è ResNet101, a cui si agganciano le teste di classificazione e di detection (ripresa da yolo). La loss definita è una combinazione di loss di classificazione e detection che in backpropagation aggiorna i pesi del backbone in base ai due task. I risultati ottenuti mostrano buone performance e miglioramenti tangibili delle metriche di classificazione del multitask a confronto con il multilabel, tra cui spicca come metrica l'F1-score che raggiunge per tutte le classi un valore maggiore di 0.90 dimostrando l'efficacia del metodo proposto.
Dalla classificazione all’identificazione basata su Landmark: un approccio Multitask multilabel per piani ecografici del cervello fetale
TIGANO, COSTANTINO
2024/2025
Abstract
In the context of fetal neurosonography, the detection of standard planes plays a fundamental role, as the analysis of these specific views enables the monitoring of fetal anatomical structures throughout gestational age and provides key measurements for the assessment of potential abnormalities. Artificial intelligence, and in particular convolutional neural networks (CNNs), is extensively applied in the acquisition of standard planes. In the literature, this task has generally been addressed as a multiclass classification problem. In this work, the problem is approached from a localization perspective, explicitly focusing on the identification of anatomical landmarks within ultrasound images. To this end, a multitask multilabel-detection framework is proposed, in which the detection branch serves as an auxiliary component to the multilabel classification task, providing additional spatial awareness during training. Building upon a baseline multilabel classification model, the proposed framework is evaluated by comparing its classification metrics with those of the multitask approach. The architecture is based on a ResNet101 backbone connected to a classification head and a detection head (adapted from yolo). The loss function combines classification and detection objectives, allowing the backbone weights to be updated jointly during backpropagation across both tasks. Experimental results demonstrate strong performance and tangible improvements in classification metrics with the multitask framework compared to the multilabel baseline. Notably, the F1-score consistently exceeds 0.90 across all classes, confirming the effectiveness and robustness of the proposed method.| File | Dimensione | Formato | |
|---|---|---|---|
| Tesi.pdf non disponibili 
										Dimensione
										1 MB
									 
										Formato
										Adobe PDF
									 | 1 MB | Adobe PDF | 
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/23246
			
		
	
	
	
			      	