Simple prostatectomy for benign prostatic hyperplasia requires identification of the cleavage plane between the adenoma and surrounding tissue, with the bladder neck and peripheral zone as anatomical landmarks. However, these structures are not always clearly distinguishable. This work develops a real-time deep learning pipeline for their segmentation to provide anatomical guidance during robotic surgery. A patient-level annotated dataset was created from endoscopic recordings using a modified SAM2-based workflow. YOLO-seg, U-Net, EndoViT-UNet, and PIDNet were evaluated, with architecture- and loss-specific modifications introduced for each; PIDNet, in particular, was extended with a pairwise boundary loss targeting the bladder neck–peripheral zone interface. PIDNet and EndoViT-UNet, selected for their complementary strengths, were then combined in a softmax ensemble. On the held-out test set, the ensemble achieved a mean IoU of 0.47 and mean Dice of 0.60 across the three foreground classes, according to the macro-metrics. The bladder neck was the most reliably segmented structure (Dice 0.67), while the adenoma remained the most challenging due to its ambiguous boundary with the peripheral zone. The final pipeline was integrated into the TilePro display and achieved over 30 FPS end-to-end through pipeline optimization. Intraoperative evaluation supported real-time usability, while highlighting uncertainty in confidently delineating the plane at the adenoma–peripheral zone interface. The results on 10 procedures support its potential as an anatomical guidance and training tool, although further improvements are needed.
La prostatectomia semplice per l'iperplasia prostatica benigna richiede l'identificazione del piano di clivaggio tra l'adenoma e il tessuto circostante, con il collo vescicale e la zona periferica come punti di riferimento anatomici. Tuttavia, queste strutture non sono sempre chiaramente distinguibili. Questo lavoro sviluppa una pipeline di deep learning in tempo reale per la loro segmentazione, al fine di fornire una guida anatomica durante la chirurgia robotica. È stato creato un dataset annotato a livello di paziente a partire da registrazioni endoscopiche, tramite un workflow modificato basato su SAM2. Sono stati valutati YOLO-seg, U-Net, EndoViT-UNet e PIDNet, ciascuno con modifiche specifiche all'architettura e alla loss; PIDNet, in particolare, è stato esteso con una pairwise boundary loss mirata all'interfaccia collo vescicale–zona periferica. PIDNet ed EndoViT-UNet, selezionati per le loro caratteristiche complementari, sono stati poi combinati in un ensemble softmax. Sul test set, l'ensemble ha raggiunto un IoU medio di 0.47 e un Dice medio di 0.60 sulle tre classi anatomiche, secondo le macro-metriche. Il collo vescicale è risultato la struttura segmentata in modo più affidabile (Dice 0.67), mentre l'adenoma è rimasto il più difficile, a causa del confine ambiguo con la zona periferica. La pipeline finale è stata integrata nel display TilePro e ha raggiunto oltre 30 FPS end-to-end grazie all'ottimizzazione della pipeline. La valutazione intraoperatoria ha confermato l'impiego in tempo reale, evidenziando al contempo un'incertezza nel delineare con sicurezza il piano a livello dell'interfaccia adenoma–zona periferica. I risultati su 10 procedure ne supportano il potenziale come strumento di guida anatomica e istruzione, sebbene siano necessari ulteriori miglioramenti.
Multi-class segmentation for real-time intraoperative surgical plane identification in simple prostatectomy
Gumiero, Chiara
2025/2026
Abstract
Simple prostatectomy for benign prostatic hyperplasia requires identification of the cleavage plane between the adenoma and surrounding tissue, with the bladder neck and peripheral zone as anatomical landmarks. However, these structures are not always clearly distinguishable. This work develops a real-time deep learning pipeline for their segmentation to provide anatomical guidance during robotic surgery. A patient-level annotated dataset was created from endoscopic recordings using a modified SAM2-based workflow. YOLO-seg, U-Net, EndoViT-UNet, and PIDNet were evaluated, with architecture- and loss-specific modifications introduced for each; PIDNet, in particular, was extended with a pairwise boundary loss targeting the bladder neck–peripheral zone interface. PIDNet and EndoViT-UNet, selected for their complementary strengths, were then combined in a softmax ensemble. On the held-out test set, the ensemble achieved a mean IoU of 0.47 and mean Dice of 0.60 across the three foreground classes, according to the macro-metrics. The bladder neck was the most reliably segmented structure (Dice 0.67), while the adenoma remained the most challenging due to its ambiguous boundary with the peripheral zone. The final pipeline was integrated into the TilePro display and achieved over 30 FPS end-to-end through pipeline optimization. Intraoperative evaluation supported real-time usability, while highlighting uncertainty in confidently delineating the plane at the adenoma–peripheral zone interface. The results on 10 procedures support its potential as an anatomical guidance and training tool, although further improvements are needed.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_10_Gumiero_ExecutiveSummary.pdf
solo utenti autorizzati a partire dal 23/09/2027
Descrizione: executive summary
Dimensione
28.09 MB
Formato
Adobe PDF
|
28.09 MB | Adobe PDF | Visualizza/Apri |
|
2026_10_Gumiero_Tesi.pdf
solo utenti autorizzati a partire dal 23/09/2027
Descrizione: Thesis
Dimensione
26.91 MB
Formato
Adobe PDF
|
26.91 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/10589/266555