Minimally invasive surgery (MIS) has revolutionized the world of surgery, by minimizing the size of the incisions, enabling a faster and less painful recovery time and, as a result, shorter hospital stays. Additionally, the introduction of robots in the operative room has shown great advantages for surgeons, giving birth to a new field called Robot-assisted minimally invasive surgery (RAMIS). In this context, Artificial Intelligence (AI) has found large application. Several algorithms have been developed to further aid the surgeon during operations, for instance by preventing certain tool movements that could cause damage to organs or vessels. To this aim, exact knowledge of the tool shape is often required, which can be automatically segmented from endoscopic images. In this thesis work, the problem of automatically segmenting surgical tools in endoscopic images is addressed. An adversarial architecture is proposed, where the generator is trained at extracting the tool mask and redrawing a new object at the same location. The generator is controlled by a discriminator that ensures that the distribution of generated images is aligned to the original one. In addition, in order to overcome the scarcity of annotated data typical of the biomedical domain, a self-supervised training approach is proposed to exploit both labelled and unlabelled data. Specifically, the generator gradually learns from unlabelled data by generating soft pseudo-labels using the network predictions. Experiments showed superior performance of the proposed method compared to the U-net architecture, which has been used as representative of state-of-the-art method. The best performance was achieved when using only 10% of the available annotated data, with a Dice Similarity Coefficient of 84.17 % with Inter Quartile Range (IQR) 14.36%. These results suggest that the proposed approach is effective when large amounts of unlabelled data combined and a low fraction of labelled data are available.
La chirurgia mininvasiva (MIS) ha rivoluzionato il mondo della chirurgia, riducendo al minimo le dimensioni delle incisioni, consentendo tempi di recupero più rapidi e meno dolorosi e, di conseguenza, degenze ospedaliere più brevi. Inoltre l’introduzione dei robot in sala operatoria ha mostrato grandi vantaggi per i chirurghi, dando vita a un nuovo campo chiamato Chirurgia Mininvasiva Assistita da Robot (RAMIS). In questo contesto, l’Intelligenza Artificiale (AI) ha trovato ampia applicazione. Sono stati sviluppati diversi algoritmi per aiutare ulteriormente il chirurgo durante le operazioni, ad esempio prevenendo determinati movimenti dello strumento che potrebbero causare danni a organi o vasi. A tal fine è spesso richiesta una conoscenza esatta della forma dello strumento, che può essere segmentata automaticamente da immagini endoscopiche. In questo lavoro di tesi viene affrontato il problema della segmentazione automatica degli strumenti chirurgici nelle immagini endoscopiche. Viene proposta un’architettura avversaria, in cui il generatore è addestrato a estrarre la maschera dello strumento e ridisegnare un nuovo oggetto nella stessa posizione. Il generatore è controllato da un discriminatore che assicura che la distribuzione delle immagini generate sia allineata a quella originale. Inoltre, al fine di superare la scarsità di dati annotati tipica del dominio biomedico, viene proposto un approccio di training auto-supervisionato per sfruttare sia dati etichettati che non etichettati. In particolare, il generatore apprende gradualmente dai dati senza etichetta generando pseudo-labels utilizzando le previsioni della rete. Gli esperimenti hanno dimostrato che le performance del metodo proposto sono superiori a quelle dell’architettura U-net, che è stata utilizzata come rappresentante dello stato dell’arte. Le migliori prestazioni sono state ottenute utilizzando solo il 10 % dei dati annotati disponibili, con un coefficiente Dice dell’84,17 % e con range interquartile (IQR) 14,36 %. Questi risultati mostrano che l’approccio proposto è efficace quando sono disponibili grandi quantità di dati non etichettati associati ad una bassa frazione di dati etichettati.
Semi-supervised segmentation of surgical instruments via pseudo-labelling and redrawing
Russo, Andrea
2020/2021
Abstract
Minimally invasive surgery (MIS) has revolutionized the world of surgery, by minimizing the size of the incisions, enabling a faster and less painful recovery time and, as a result, shorter hospital stays. Additionally, the introduction of robots in the operative room has shown great advantages for surgeons, giving birth to a new field called Robot-assisted minimally invasive surgery (RAMIS). In this context, Artificial Intelligence (AI) has found large application. Several algorithms have been developed to further aid the surgeon during operations, for instance by preventing certain tool movements that could cause damage to organs or vessels. To this aim, exact knowledge of the tool shape is often required, which can be automatically segmented from endoscopic images. In this thesis work, the problem of automatically segmenting surgical tools in endoscopic images is addressed. An adversarial architecture is proposed, where the generator is trained at extracting the tool mask and redrawing a new object at the same location. The generator is controlled by a discriminator that ensures that the distribution of generated images is aligned to the original one. In addition, in order to overcome the scarcity of annotated data typical of the biomedical domain, a self-supervised training approach is proposed to exploit both labelled and unlabelled data. Specifically, the generator gradually learns from unlabelled data by generating soft pseudo-labels using the network predictions. Experiments showed superior performance of the proposed method compared to the U-net architecture, which has been used as representative of state-of-the-art method. The best performance was achieved when using only 10% of the available annotated data, with a Dice Similarity Coefficient of 84.17 % with Inter Quartile Range (IQR) 14.36%. These results suggest that the proposed approach is effective when large amounts of unlabelled data combined and a low fraction of labelled data are available.| File | Dimensione | Formato | |
|---|---|---|---|
|
2021_04_Russo.pdf
non accessibile
Descrizione: testo della tesi
Dimensione
3.54 MB
Formato
Adobe PDF
|
3.54 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/10589/175036