Globally, colorectal cancer accounts for about 10% of all cancer diagnoses, making it one of the leading causes of cancer-related mortality. Its prevention primarily relies on the early detection of polyps in the colon and rectum through colonoscopy. However, this procedure has certain limitations, as it is prone to human errors, and its accuracy depends on the operator’s level of experience. For these reasons, in recent years, there has been a growing interest in the development of deep learning-based systems to assist healthcare professionals in the automatic detection of colorectal polyps. The integration of these technologies into the diagnostic field aims to improve the accuracy of lesion detection by optimizing the detection rate and, consequently, reducing potential human errors. However, even these deep learning-based models are not infallible, and their effectiveness is influenced by a variety of intrinsic and extrinsic factors. This thesis aims to investigate the application of the RT-DETR model for the automatic detection of colorectal polyps, analyzing its performance on the "REAL Colon" dataset. This dataset represents a collection of full colonoscopy videos recorded from start to finish without pauses or interruptions; moreover, it includes different endoscopic practices from various geographical regions, thus ensuring a wide variability of the data. The goal of the study is to assess the effectiveness of individual data augmentation techniques in improving the model’s detection performance, identifying the most effective ones to optimize the accuracy and reliability of automated lesion detection. The results obtained from this analysis provide valuable insights for optimizing diagnostic support systems in gastroenterology, highlighting which data augmentation techniques are most suitable and how artificial intelligence can positively influence the model’s performance in the prevention and early diagnosis of colorectal cancer.
A livello globale, il tumore colorettale costituisce circa il 10% di tutte le diagnosi oncologiche diventando così una delle principali cause di mortalità per cancro. La sua prevenzione si basa principalmente sull'individuazione precoce dei polipi nel colon e nel retto attraverso la colonscopia. Tuttavia, questa procedura presenta alcune limitazioni, in quanto può essere soggetta a errori umani e la sua accuratezza dipende dal livello di esperienza dell'operatore. Per queste motivazioni, negli ultimi anni si è registrato un crescente interesse nello sviluppo di sistemi basati sul deep learning al fine di supportare ed aiutare i professionisti sanitari nella rilevazione automatica dei polipi colorettali. L’integrazione di queste tecnologie in ambito diagnostico, dunque, ha l’obiettivo di migliorare l’accuratezza nell’individuazione delle lesioni ottimizzando il tasso di rilevamento e riducendo, di conseguenza, i potenziali errori umani. Tuttavia, anche questi modelli basati sul deep learning non sono infallibili e la loro efficacia è influenzata da una serie di fattori intrinseci ed estrinseci al modello stesso. Questa tesi si è posta l’obbiettivo di indagare l’applicazione del modello RT-DETR per il rilevamento automatico dei polipi colorettali, analizzandone le performance sul dataset “REAL Colon”. Questo dataset rappresenta una collezione di video integrali di colonscopie registrate dall’inizio alla fine senza pause o interruzioni; inoltre, integra pratiche endoscopiche diverse tra loro e provenienti da regioni geografiche differenti, garantendo così un’ampia variabilità dei dati. L’obiettivo dello studio è dunque quello di valutare l’efficacia delle singole tecniche di data augmentation nel migliorare le prestazioni del modello di rilevamento, individuando quelle più efficaci per ottimizzare l’accuratezza e l’affidabilità della rilevazione automatizzata delle lesioni. I risultati ottenuti da questa analisi forniscono utili spunti per ottimizzare i sistemi di supporto alla diagnosi in ambito gastroenterologico, evidenziando quali tecniche di data augmentation risultino più idonee e come l’intelligenza artificiale possa influenzare positivamente le prestazioni del modello nella prevenzione e della diagnosi precoce del tumore colorettale.
Confronto delle tecniche di data augmentation applicate al rilevamento dei polipi in colonscopia
FORNASIERO, CECILIA
2024/2025
Abstract
Globally, colorectal cancer accounts for about 10% of all cancer diagnoses, making it one of the leading causes of cancer-related mortality. Its prevention primarily relies on the early detection of polyps in the colon and rectum through colonoscopy. However, this procedure has certain limitations, as it is prone to human errors, and its accuracy depends on the operator’s level of experience. For these reasons, in recent years, there has been a growing interest in the development of deep learning-based systems to assist healthcare professionals in the automatic detection of colorectal polyps. The integration of these technologies into the diagnostic field aims to improve the accuracy of lesion detection by optimizing the detection rate and, consequently, reducing potential human errors. However, even these deep learning-based models are not infallible, and their effectiveness is influenced by a variety of intrinsic and extrinsic factors. This thesis aims to investigate the application of the RT-DETR model for the automatic detection of colorectal polyps, analyzing its performance on the "REAL Colon" dataset. This dataset represents a collection of full colonoscopy videos recorded from start to finish without pauses or interruptions; moreover, it includes different endoscopic practices from various geographical regions, thus ensuring a wide variability of the data. The goal of the study is to assess the effectiveness of individual data augmentation techniques in improving the model’s detection performance, identifying the most effective ones to optimize the accuracy and reliability of automated lesion detection. The results obtained from this analysis provide valuable insights for optimizing diagnostic support systems in gastroenterology, highlighting which data augmentation techniques are most suitable and how artificial intelligence can positively influence the model’s performance in the prevention and early diagnosis of colorectal cancer.File | Dimensione | Formato | |
---|---|---|---|
Fornasiero_Cecilia.pdf
accesso riservato
Dimensione
3.61 MB
Formato
Adobe PDF
|
3.61 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/83030