Digital pathology has enabled the integration of computational methods into clinical work- flows through the analysis of Whole Slide Images. However, the presence of preparation and acquisition artifacts, such as tissue folds or external objects, can negatively affect both human diagnosis and the performance of computational pathology systems. This research aims to develop a robust, computationally efficient quality control pipeline by adapting Vision Transformers for automated artifact detection. Furthermore, it seeks to formally model and structurally integrate the detection of scanner-specific stitching ar- tifacts into a unified deep learning architecture. A novel multiscale Vision Transformer architecture is proposed to effectively capture both local and global features, leveraging a two-branch framework that processes concentrically cropped image patches at differ- ent magnifications. To improve training procedure that is tainted by the scarcity of annotated data, a knowledge distillation strategy was adopted using artifact masks gen- erated by a Convolutional Neural Network teacher model. Results demonstrate that the proposed model significantly outperformed both the baseline teacher model and a single- scale Vision Transformer architecture in terms of segmentation accuracy. Additionally, the structural integration of the stitching artifact detection module successfully identified scanner-specific anomalies without compromising the performance of the overall quality control pipeline and with minimal computational overhead. Automated, early artifact detection streamlines the diagnostic process, reducing the time and resources spent on manual review and re-scanning, ultimately improving patient outcomes.
La patologia digitale ha permesso l’integrazione di metodi computazionali nei flussi di lavoro clinici attraverso l’analisi di Whole Slide Images (immagini a intero vetrino). Tut- tavia, la presenza di artefatti di preparazione e acquisizione, come pieghe del tessuto o oggetti esterni, può influire negativamente sia sulla diagnosi umana che sulle prestazioni dei sistemi di computational pathology. Questa ricerca mira a sviluppare una pipeline di controllo qualità robusta ed efficiente dal punto di vista computazionale, adattando un’architettura basata su Vision Transformer per il rilevamento automatizzato degli arte- fatti. Inoltre, lo studio cerca di modellare formalmente e integrare il rilevamento di arte- fatti di stitching specifici dello scanner all’interno di un’architettura unificata. Viene pro- posta una nuova architettura Vision Transformer multiscala per catturare efficacemente sia le feature locali che quelle globali, sfruttando un modello a due rami che elabora ri- quadri di immagini ritagliate concentricamente a diversi ingrandimenti. Per migliorare la procedura di training, penalizzata dalla scarsità di dati annotati, è stata adottata una strategia di knowledge distillation utilizzando maschere di artefatti generate da un modello teacher basato su una Convolutional Neural Network. I risultati dimostrano che il modello proposto supera significativamente sia il modello teacher di baseline, sia un’architettura Vision Transformer a singola scala in termini di accuratezza della segmentazione. Inoltre, l’integrazione del modulo di rilevamento degli artefatti di stitching ha identificato con successo anomalie specifiche dello scanner senza compromettere le prestazioni dell’intera pipeline di controllo qualità e con un overhead computazionale minimo. Il rilevamento au- tomatizzato e precoce degli artefatti ottimizza il processo diagnostico, riducendo il tempo e le risorse impiegate per la revisione manuale e il re-scanning, migliorando in ultima analisi gli esiti per i pazienti.
Efficient multiscale deep learning for artifact detection in digital pathology
FACCIOLI, ANDREA
2025/2026
Abstract
Digital pathology has enabled the integration of computational methods into clinical work- flows through the analysis of Whole Slide Images. However, the presence of preparation and acquisition artifacts, such as tissue folds or external objects, can negatively affect both human diagnosis and the performance of computational pathology systems. This research aims to develop a robust, computationally efficient quality control pipeline by adapting Vision Transformers for automated artifact detection. Furthermore, it seeks to formally model and structurally integrate the detection of scanner-specific stitching ar- tifacts into a unified deep learning architecture. A novel multiscale Vision Transformer architecture is proposed to effectively capture both local and global features, leveraging a two-branch framework that processes concentrically cropped image patches at differ- ent magnifications. To improve training procedure that is tainted by the scarcity of annotated data, a knowledge distillation strategy was adopted using artifact masks gen- erated by a Convolutional Neural Network teacher model. Results demonstrate that the proposed model significantly outperformed both the baseline teacher model and a single- scale Vision Transformer architecture in terms of segmentation accuracy. Additionally, the structural integration of the stitching artifact detection module successfully identified scanner-specific anomalies without compromising the performance of the overall quality control pipeline and with minimal computational overhead. Automated, early artifact detection streamlines the diagnostic process, reducing the time and resources spent on manual review and re-scanning, ultimately improving patient outcomes.| File | Dimensione | Formato | |
|---|---|---|---|
|
AndreaF_Thesis_Polimi.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
56.77 MB
Formato
Adobe PDF
|
56.77 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/10589/260181