This thesis explores the integration of artificial intelligence (AI) in Otolaryngology – Head and Neck Surgery, focusing on advancements in computer vision for endoscopy and surgical procedures. It begins with a comprehensive review of AI and computer vision advancements in this field, identifying areas for further exploration. The primary aim was to develop a computer vision system for endoscopy analysis. The research involved designing tools for detecting and segmenting neoplasms in the upper aerodigestive tract (UADT) and assessing vocal fold motility, crucial in laryngeal cancer staging. Further, the study delves into the potential of vision foundation models, like vision transformers trained via self-supervision, to reduce the need for expert annotations, particularly beneficial in fields with limited cases. Additionally, the research includes the development of a web application for enhancing and speeding up the annotation process in UADT endoscopy, under the umbrella of Machine Learning Operations (MLOps). The thesis covers various phases of research, starting with defining the conceptual framework and methodology, termed "Videomics". It includes a literature review on AI in clinical endoscopy, focusing on Narrow Band Imaging (NBI) and convolutional neural networks (CNNs). The research progresses through different stages, from quality assessment of endoscopic images to in-depth characterization of neoplastic lesions. It also addresses the need for standards in medical computer vision study reporting and evaluates the application of AI in dynamic vision scenarios like vocal fold motility. A significant part of the research investigates the use of "general purpose" vision algorithms and the commoditization of machine learning algorithms, using nasal polyps and oropharyngeal cancer as case studies. Finally, the thesis discusses the development of ENDO-CLOUD, a cloud-based system for videolaryngoscopy, highlighting the challenges and solutions in data management and the large-scale deployment of AI models in medical imaging.
Questa tesi esplora l'integrazione dell'intelligenza artificiale (IA) in Otorinolaringoiatria – Chirurgia di Testa e Collo, concentrandosi sui progressi della computer vision per l’endoscopia e le procedure chirurgiche. La ricerca inizia con una revisione completa dello stato dell’arte dell'IA e della computer vision in questo campo, identificando aree per ulteriori sviluppi. L'obiettivo principale è stato quello di sviluppare un sistema di computer vision per l'analisi di immagini e video endoscopici. La ricerca ha coinvolto la progettazione di strumenti per la rilevazione e segmentazione di neoplasie nelle vie aerodigestive superiori (VADS) e la valutazione della motilità delle corde vocali, cruciale nella stadiazione del carcinoma laringeo. Inoltre, lo studio si è focalizzato sul potenziale dei foundation vision models, vision transformers basati su self-supervised learning, per ridurre la necessità di annotazione da parte di esperti, approccio particolarmente vantaggioso in campi con dati limitati. Inoltre, la ricerca ha incluso lo sviluppo di un'applicazione web per migliorare e velocizzare il processo di annotazione in endoscopia delle VADS, nell’ambito generale delle tecniche di MLOps. La tesi copre varie fasi della ricerca, a partire dalla definizione del quadro concettuale e della metodologia, denominata "Videomics". Include una revisione della letteratura sull'IA in endoscopia clinica, focalizzata sulla Narrow Band Imaging (NBI) e sulle reti neurali convoluzionali (CNN). Lo studio progredisce attraverso diverse fasi, dalla valutazione della qualità delle immagini endoscopiche alla caratterizzazione approfondita delle lesioni neoplastiche. Si affronta anche la necessità di standard nel reporting degli studi di computer vision in ambito medico e si valuta l'applicazione dell'IA in setting dinamici come la motilità delle corde vocali. Una parte significativa della ricerca indaga l'uso di algoritmi di computer vision generalizzati (“foundation models”) e la “commoditization” degli algoritmi di machine learning, utilizzando polipi nasali e il carcinoma orofaringeo come casi studio. Infine, la tesi discute lo sviluppo di ENDO-CLOUD, un sistema basato su cloud per l’analisi della videolaringoscopia, evidenziando le sfide e le soluzioni nella gestione dei dati e l’utilizzo su larga scala di modelli di IA nell'imaging medico.
SCALING ARTIFICIAL INTELLIGENCE IN ENDOSCOPY: FROM MODEL DEVELOPMENT TO MACHINE LEARNING OPERATIONS FRAMEWORKS / Paderno, Alberto. - (2024 Apr 11).
SCALING ARTIFICIAL INTELLIGENCE IN ENDOSCOPY: FROM MODEL DEVELOPMENT TO MACHINE LEARNING OPERATIONS FRAMEWORKS
PADERNO, ALBERTO
2024-04-11
Abstract
This thesis explores the integration of artificial intelligence (AI) in Otolaryngology – Head and Neck Surgery, focusing on advancements in computer vision for endoscopy and surgical procedures. It begins with a comprehensive review of AI and computer vision advancements in this field, identifying areas for further exploration. The primary aim was to develop a computer vision system for endoscopy analysis. The research involved designing tools for detecting and segmenting neoplasms in the upper aerodigestive tract (UADT) and assessing vocal fold motility, crucial in laryngeal cancer staging. Further, the study delves into the potential of vision foundation models, like vision transformers trained via self-supervision, to reduce the need for expert annotations, particularly beneficial in fields with limited cases. Additionally, the research includes the development of a web application for enhancing and speeding up the annotation process in UADT endoscopy, under the umbrella of Machine Learning Operations (MLOps). The thesis covers various phases of research, starting with defining the conceptual framework and methodology, termed "Videomics". It includes a literature review on AI in clinical endoscopy, focusing on Narrow Band Imaging (NBI) and convolutional neural networks (CNNs). The research progresses through different stages, from quality assessment of endoscopic images to in-depth characterization of neoplastic lesions. It also addresses the need for standards in medical computer vision study reporting and evaluates the application of AI in dynamic vision scenarios like vocal fold motility. A significant part of the research investigates the use of "general purpose" vision algorithms and the commoditization of machine learning algorithms, using nasal polyps and oropharyngeal cancer as case studies. Finally, the thesis discusses the development of ENDO-CLOUD, a cloud-based system for videolaryngoscopy, highlighting the challenges and solutions in data management and the large-scale deployment of AI models in medical imaging.File | Dimensione | Formato | |
---|---|---|---|
Tesi PhD Paderno Corretto.pdf
accesso aperto
Descrizione: Tesi
Tipologia:
Tesi di dottorato
Dimensione
18.43 MB
Formato
Adobe PDF
|
18.43 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.