Colorectal cancer (CRC) is one of the leading causes of death worldwide and continues to pose a critical public health challenge, demanding precise early detection and intervention. Colonoscopy, the diagnostic examination aimed at exploring the inner walls of the colon to discover any tumour masses, is an effective method to decrease mortality incidence. Emerging techniques, such as advanced image analysis driven by neural networks, hold promise for accurate diagnosis. However, studies have reported that, for various reasons, a certain percentage of polyps are not correctly detected during colonoscopy. One of the most important is the dependency on pixel-level annotations, which requires a lot of computational resources, making necessary innovative solutions. This thesis introduces strategies for improving polyp identification. For this purpose, the main techniques involve the so-called Explainable AI tools for analyzing saliency maps and activation maps, through several state-of-the-art visual saliency detection algorithms and Gradient-weighted Class Activation Mapping (Grad-CAM). In addition, a neural network for segmentation with DeepLabV3+ architecture is used, in which bounding boxes are provided on the training images, within a weakly supervised framework.
Il cancro del colon-retto (CRC) è una delle principali cause di morte a livello mondiale e continua a rappresentare una sfida critica per la salute pubblica, richiedendo una precisa e tempestiva diagnosi e un intervento mirato. La colonscopia, ovvero l'esame diagnostico volto a esplorare le pareti interne del colon per scoprire eventuali masse tumorali, ha dimostrato essere un metodo efficace per ridurre l'incidenza di mortalità. Le tecniche emergenti, come l'analisi avanzata delle immagini tramite reti neurali, sono promettenti per una diagnosi accurata. Tuttavia, alcuni studi hanno riportato che, per varie ragioni, una certa percentuale di polipi non viene rilevata correttamente durante la colonscopia. Una delle più importanti è la dipendenza dalle annotazioni a livello di pixel, che richiede molte risorse computazionali; per questo si rendono necessarie soluzioni innovative. Questa tesi introduce alcune strategie per migliorare l'identificazione dei polipi. A tal fine, le tecniche principali utilizzate coinvolgono i cosiddetti metodi di Explainable AI per l'analisi delle mappe di salienza e di attivazione, attraverso diversi algoritmi di rilevamento della salienza visiva e la Gradient-weighted Class Activation Mapping (Grad-CAM). Inoltre, viene utilizzata una rete neurale per la segmentazione con architettura DeepLabV3+, in cui vengono fornite le bounding box sulle immagini di addestramento, in un contesto debolmente supervisionato.
Segmentazione debolmente supervisionata di polipi su immagini di colonscopia
RAMPON, RICCARDO
2022/2023
Abstract
Colorectal cancer (CRC) is one of the leading causes of death worldwide and continues to pose a critical public health challenge, demanding precise early detection and intervention. Colonoscopy, the diagnostic examination aimed at exploring the inner walls of the colon to discover any tumour masses, is an effective method to decrease mortality incidence. Emerging techniques, such as advanced image analysis driven by neural networks, hold promise for accurate diagnosis. However, studies have reported that, for various reasons, a certain percentage of polyps are not correctly detected during colonoscopy. One of the most important is the dependency on pixel-level annotations, which requires a lot of computational resources, making necessary innovative solutions. This thesis introduces strategies for improving polyp identification. For this purpose, the main techniques involve the so-called Explainable AI tools for analyzing saliency maps and activation maps, through several state-of-the-art visual saliency detection algorithms and Gradient-weighted Class Activation Mapping (Grad-CAM). In addition, a neural network for segmentation with DeepLabV3+ architecture is used, in which bounding boxes are provided on the training images, within a weakly supervised framework.File | Dimensione | Formato | |
---|---|---|---|
Rampon_Riccardo.pdf
accesso aperto
Dimensione
6.98 MB
Formato
Adobe PDF
|
6.98 MB | Adobe PDF | Visualizza/Apri |
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/20.500.12608/55808