The increase in the complexity of the techniques required to detect and perform diagnosis “in vivo” with high quality stan- dards and the subsequent need of highly skilled endoscopists has boosted the interest in the implementation of Artificial Intelligence algorithms to support the practitioner during Colorectal Cancer Screening colonoscopies. Even though Deep Learning Convolutional Neural Networks applied to the detection and classification of superficial neoplastic lesions have shown promising results, they still have several limitations mainly related to the lack of large enough real world datasets. In fact, most of currently available data is confined to research purposes and do not reflect the variety of conditions that the clinicians face in the daily practice. Therefore, this thesis aims at building a novel real-world polyp multi-label classification database and to validate a scalable data collection pipeline that can be implemented in the Endoscopic Operative Units. The second objective is to use the just created dataset to investigate the feasibility of extending current off-the-shelf state-of-the-art ResNet-50 and Capsule Network models to the Paris and Nice classification taxonomies adopted by endoscopists to assess the depth of invasion and severity of the lesions. A preliminary dataset of 525 polyp sequences representing 191 unique polyps was collected from 100 colonoscopies for a total of 26’000 frames, in Fondazione Poliambulanza Hospital Brescia (IT). This private dataset is to be considered one of the biggest for polyp classification tasks and spans across more than one year of clinical practice. Among the trained models the best performances have been obtained by the ResNet-50 which obtained 56% ± 12% accuracy, 57% ± 12% precision and 0.56% ± 10% recall during 5 fold cross validation. All the code is available at the Github repository of the project. ( https://github.com/Stefano97/polyp-classification-poliambulanza.git )
L’aumento in complessità delle tecniche richieste ad individuare ed eseguire una diagnosi "in vivo" con standard di alta qualità e la conseguente necessità di endoscopisti esperti, ha incentivato l’interesse nell’implementazione di algoritmi di Intelligenza Artificiale a supporto del medico durante colonscopie di screening per cancro al colon-retto. Nonostante reti neurali convoluzionali di Deep Learning applicate alla detezione e classificazione di lesioni neoplastiche superficiali ha dimostrato risultati promettenti, sono presenti molte limitazioni associate alla mancanza di dataset provenienti dalla pratica clinica di sufficienti dimensioni. Infatti, la maggior parte dei dataset attualmente disponibili sono limitati alla ricerca e non riflettono la varietà di condizioni che il clinico affronta nella pratica quotidiana. Per queste ragioni, questa tesi ha come obiettivo la costruzione di un nuovo dataset per la classificazione multi-classe di polipi e contestualmente la validazione del metodo di raccolta dati che possa essere esteso ed introdotto nelle Unità Operative di Endoscopia. Il secondo obiettivo è quello di investigare la fattibilità di estendere i modelli ResNet-50 e Capsule Network considerati stato dell’arte alle tassonomie di Parigi e Nice utilizzando il dataset appena creato. Il dataset ancora nella sua fase preliminare contiene 525 sequenze di polipi che rappresentano 191 polipi unici identificati in 100 colonscopie per un totale di 26’000 frames, presso l’ospedale di Fondazione Poliambualanza a Brescia (Italia). Questo dataset privato è da considerarsi il più grande dataset per la classificazione di lesioni superficiali neoplastiche e si sviluppa su più di un anno di attività clinica. Fra i modelli allenati quello con le migliori performance è ResNet-50 che ha ottenuto 56% ± 12% di accuratezza, 57% ± 12% di precisione e 0.56% ± 10% di recall durante la cross-validazione a 5 fold. Il codice è disponibile alla GitHub repository del progetto. ( https://github.com/Stefano97/polyp-classification-poliambulanza.git )
Construction of a novel database and evaluation of artificial neural networks for endoscopic classification of superficial neoplastic lesions
Magni, Stefano
2020/2021
Abstract
The increase in the complexity of the techniques required to detect and perform diagnosis “in vivo” with high quality stan- dards and the subsequent need of highly skilled endoscopists has boosted the interest in the implementation of Artificial Intelligence algorithms to support the practitioner during Colorectal Cancer Screening colonoscopies. Even though Deep Learning Convolutional Neural Networks applied to the detection and classification of superficial neoplastic lesions have shown promising results, they still have several limitations mainly related to the lack of large enough real world datasets. In fact, most of currently available data is confined to research purposes and do not reflect the variety of conditions that the clinicians face in the daily practice. Therefore, this thesis aims at building a novel real-world polyp multi-label classification database and to validate a scalable data collection pipeline that can be implemented in the Endoscopic Operative Units. The second objective is to use the just created dataset to investigate the feasibility of extending current off-the-shelf state-of-the-art ResNet-50 and Capsule Network models to the Paris and Nice classification taxonomies adopted by endoscopists to assess the depth of invasion and severity of the lesions. A preliminary dataset of 525 polyp sequences representing 191 unique polyps was collected from 100 colonoscopies for a total of 26’000 frames, in Fondazione Poliambulanza Hospital Brescia (IT). This private dataset is to be considered one of the biggest for polyp classification tasks and spans across more than one year of clinical practice. Among the trained models the best performances have been obtained by the ResNet-50 which obtained 56% ± 12% accuracy, 57% ± 12% precision and 0.56% ± 10% recall during 5 fold cross validation. All the code is available at the Github repository of the project. ( https://github.com/Stefano97/polyp-classification-poliambulanza.git )| File | Dimensione | Formato | |
|---|---|---|---|
|
2021_12_Magni.pdf
non accessibile
Dimensione
16.61 MB
Formato
Adobe PDF
|
16.61 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/10589/182892