Background: The diagnosis of Hirschsprung disease (HD) is definitively confirmed through the histopathological analysis of rectal biopsies. However, the manual examination of histological slides is a labor-intensive process prone to errors, highlighting the need to utilize Artificial Intelligence (AI) as a decision- support tool. The application of Machine Learning (ML) in digital pathology is often limited by the scarcity of large datasets of high-quality images. Aim: This study proposes an innovative approach whose aim is to create an optimized dataset of real histological images to be used as the basis for establishing a generative model of similar synthetic images in order to increase the number of available images and indirectly improve the diagnostic performance of ML models in Hirschsprung's disease. Materials and Methods: Formalin-fixed samples were manually annotated to highlight ganglion cells and nerve structures. Subsequently, standard preprocessing techniques such as spatial filtering, color normalization, and segmentation were applied to improve image quality and create masks. A denoising diffusion-based generative model was then trained with the objective of synthesizing images containing ganglion cells. Results: We reviewed, selected, and digitized 108 slides from patients who underwent surgery between January 2010 and January 2022. These images were then annotated and segmented to train the Generative AI. The dataset was expanded through augmentation, which were divided into smaller sections corresponding to the input size of the network. The generative network produced novel images containing ganglion cells. Conclusion: The integration of synthetic and real images, coupled with optimized preprocessing and segmentation, represents a potential strategy to improve the model's performance in detecting the diagnostic features of HD. This study highlights the potential of using synthetic histological images in training automated systems with the goal of accelerating analysis in HD diagnosis and reducing the burden of manual evaluation.
Background: La diagnosi di malattia di Hirschsprung (HD) viene confermata attraverso l'analisi istopatologica delle biopsie rettali. Tuttavia, l'esame manuale dei vetrini istologici è un processo laborioso e soggetto ad errori, il che evidenzia la necessità di utilizzare l'Intelligenza Artificiale (AI) come strumento di supporto decisionale. L'applicazione del Machine Learning (ML) nella patologia digitale è spesso limitata dalla ridotta disponibilità di ampi set di immagini di alta qualità. Scopo: Questo studio propone un approccio innovativo il cui scopo è la creazione di un dataset ottimizzato di immagini istologiche reali da usare come base per la definizione di un modello generativo di immagini sintetiche analoghe al fine di aumentare il numero di immagini disponibili e migliorare indirettamente le prestazioni diagnostiche dei modelli di ML nella malattia di Hirschsprung. Materiali e metodi: I campioni fissati in formalina sono stati annotati manualmente evidenziando le cellule gangliari e le strutture nervose. Successivamente, sono state applicate tecniche di preprocessing standard, come lo spatial filtering, la normalizzazione dei colori e la segmentazione, per migliorare la qualità delle immagini e creare le maschere. Successivamente, è stato addestrato un modello generativo basato sulla diffusione denoising con l'obiettivo di sintetizzare immagini contenenti cellule gangliari. Risultati: Abbiamo visionato, selezionato e digitalizzato 108 slides provenienti da pazienti sottoposti ad intervento tra gennaio 2010 e gennaio 2022. Successivamente queste immagini sono state annotate e segmentate per allenare la Generative AI. Il dataset è stato ampliato, grazie all’augumentation, ed è stato suddiviso in piccole sezioni che corrispondono alle dimensioni di ingresso della rete. La rete generativa ha prodotto immagini inedite contenenti cellule gangliari. Conclusione: L’integrazione di immagini sintetiche e reali, insieme a un’ottimizzazione del preprocessing e della segmentazione, rappresenta una possibile strategia per migliorare le prestazioni del modello nell’identificazione delle caratteristiche diagnostiche di HD. Questo studio evidenzia il potenziale dell’uso combinato di immagini istologiche sintetiche nel training dei sistemi automatici con l’obiettivo di accelerare l’analisi nella diagnosi di HD e ridurre il carico della valutazione manuale.
L’intelligenza artificiale generativa come supporto per la diagnosi istopatologica della malattia di Hirschsprung
SEGAT, FRANCESCA
2024/2025
Abstract
Background: The diagnosis of Hirschsprung disease (HD) is definitively confirmed through the histopathological analysis of rectal biopsies. However, the manual examination of histological slides is a labor-intensive process prone to errors, highlighting the need to utilize Artificial Intelligence (AI) as a decision- support tool. The application of Machine Learning (ML) in digital pathology is often limited by the scarcity of large datasets of high-quality images. Aim: This study proposes an innovative approach whose aim is to create an optimized dataset of real histological images to be used as the basis for establishing a generative model of similar synthetic images in order to increase the number of available images and indirectly improve the diagnostic performance of ML models in Hirschsprung's disease. Materials and Methods: Formalin-fixed samples were manually annotated to highlight ganglion cells and nerve structures. Subsequently, standard preprocessing techniques such as spatial filtering, color normalization, and segmentation were applied to improve image quality and create masks. A denoising diffusion-based generative model was then trained with the objective of synthesizing images containing ganglion cells. Results: We reviewed, selected, and digitized 108 slides from patients who underwent surgery between January 2010 and January 2022. These images were then annotated and segmented to train the Generative AI. The dataset was expanded through augmentation, which were divided into smaller sections corresponding to the input size of the network. The generative network produced novel images containing ganglion cells. Conclusion: The integration of synthetic and real images, coupled with optimized preprocessing and segmentation, represents a potential strategy to improve the model's performance in detecting the diagnostic features of HD. This study highlights the potential of using synthetic histological images in training automated systems with the goal of accelerating analysis in HD diagnosis and reducing the burden of manual evaluation.File | Dimensione | Formato | |
---|---|---|---|
tesi Hirschsprung definitiva pdf_a.pdf
accesso aperto
Dimensione
5.03 MB
Formato
Adobe PDF
|
5.03 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/20.500.12608/82871