Disc herniation is a degenerative disorder of the intervertebral disc that affects a significant portion of the worldwide population. One of the most commonly adopted surgical treatments is microdiscectomy, a less invasive alternative to standard open discectomy. This procedure could benefit from robotic assistance; however, currently available robotic surgical systems are not fully suitable for this type of intervention, mainly due to the lack of robust computer vision systems capable of understanding the surgical scene and supporting intra-operative trajectory adaptation. This thesis proposes a computer-vision-based approach intended as a first step toward robot-assisted microdiscectomy. The system relies on a Vision Transformer (ViT)-based segmentation model designed to identify relevant anatomical structures and surgical instruments within the operative field. To date, surgical image segmentation research has mainly focused on endoscopic procedures, while open surgical scenarios have received less attention. Moreover, previous works have often relied on binary segmentation or prompt-based models, which are unsuitable for real-time intra-operative applications. The proposed architecture employs a DINO-pretrained backbone for feature extraction, coupled with a lightweight Feature Fusion Unit (FFU) that generates the final segmentation masks. Due to the limited availability of exoscopic surgical datasets, a dedicated dataset was also developed and annotated. During training, most backbone parameters are kept frozen, while the final layers and the FFU are optimized using a combination of Cross-Entropy and Dice losses with equal weighting. A validation strategy designed to prevent data leakage between training and validation sets was adopted to reliably assess model generalization. The proposed architecture achieves an overall mIoU of 0.5615 and a Dice score of 0.6092, while reaching an inference speed of 97.77 frames per second, outperforming established segmentation models such as YOLO, DeepLabV3+, and SegFormer. Overall, the proposed solution provides accurate real-time segmentation of open surgical scenes and represents a first step toward computer-vision-assisted and robot-assisted microdiscectomy
L’ernia del disco è una malattia degenerativa che colpisce i dischi intervertebrali e affligge una porzione sostanziosa della popolazione mondiale. Tra le pratiche chirurgiche più adottate per trattarla c’è la microdiscectomia, procedura che rappresenta un’alternativa meno invasiva della discectomia standard. Sebbene potrebbe beneficiare dell’introduzione dell’assistenza robotica, i sistemi di questo tipo diponibili oggi non hanno sistemi di computer vision (CV) abbastanza competenti nell’analisi di immagini chirurgiche. Questo lavoro si propone di sviluppare un modello di deep learning (DL) per muovere un primo passo verso il superamento di questo limite, appoggiandosi ad un modello di segmentazione basato su Vision-Transformer (ViT), per ottenere la segmentazione, nelle immagini, di strutture anatomiche e strumenti chirurgici di interesse per la procedura. Ad oggi la segmentazione nell’ambito chirurgico si è concentrata prevalentemente su operazioni endoscopiche, tralasciando le procedure "aperte" convnzionali di cui è parte anche la microdiscectomia. Per di più molti modelli sono basati sull’utilizzo di prompt o su segmentazione binaria, diventando quindi una opzione non viabile per applicazioni real-time. L’architettura proposta si basa sull’accopiamento di un backbone preaddestrato DINO, per estrazione delle features, ed una Feature Fusion Unit (FFU) con un basso numero di parametri, per la generazione della mask finale. Viene anche fornito un dataset specifico, basato su microdischectomia, per l’addestramento, che avviene mantenendo "freezata" gran parte del backbone mentre, il suo ultimo layer e l’FFU, sono ottimizzati tramite combinazione di Cross-Entropy e Dice Loss. Viene adottato una strategia di validazione atta a scongiurare data leakage, per analizzare nel migliore dei modi le capacità di generalizazione del modello. L’architettura proposta arriva a raggiungere un mIoU di 0.5615 ed un Dice di 0.6092, mantenendo una velocità di inferenza media di 97.77 fps, superando modelli ben consolidati in letteratura come Yolov11, DeepLabV3+ e SegFormer. Nel complesso, la soluzione proposta consente di ottenere una segmentazione accurata ed in tempo reale di immagini chirurgiche e rappresenta un primo passo verso la microdiscectomia assistita dalla computer vision e dalla robotica.
ExoDINO : real-time surgical scene instance segmentation for open microdiscectomy
Pachioli, Domenico
2025/2026
Abstract
Disc herniation is a degenerative disorder of the intervertebral disc that affects a significant portion of the worldwide population. One of the most commonly adopted surgical treatments is microdiscectomy, a less invasive alternative to standard open discectomy. This procedure could benefit from robotic assistance; however, currently available robotic surgical systems are not fully suitable for this type of intervention, mainly due to the lack of robust computer vision systems capable of understanding the surgical scene and supporting intra-operative trajectory adaptation. This thesis proposes a computer-vision-based approach intended as a first step toward robot-assisted microdiscectomy. The system relies on a Vision Transformer (ViT)-based segmentation model designed to identify relevant anatomical structures and surgical instruments within the operative field. To date, surgical image segmentation research has mainly focused on endoscopic procedures, while open surgical scenarios have received less attention. Moreover, previous works have often relied on binary segmentation or prompt-based models, which are unsuitable for real-time intra-operative applications. The proposed architecture employs a DINO-pretrained backbone for feature extraction, coupled with a lightweight Feature Fusion Unit (FFU) that generates the final segmentation masks. Due to the limited availability of exoscopic surgical datasets, a dedicated dataset was also developed and annotated. During training, most backbone parameters are kept frozen, while the final layers and the FFU are optimized using a combination of Cross-Entropy and Dice losses with equal weighting. A validation strategy designed to prevent data leakage between training and validation sets was adopted to reliably assess model generalization. The proposed architecture achieves an overall mIoU of 0.5615 and a Dice score of 0.6092, while reaching an inference speed of 97.77 frames per second, outperforming established segmentation models such as YOLO, DeepLabV3+, and SegFormer. Overall, the proposed solution provides accurate real-time segmentation of open surgical scenes and represents a first step toward computer-vision-assisted and robot-assisted microdiscectomy| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_10_Pachioli_Tesi.pdf
accessibile in internet per tutti a partire dal 01/10/2029
Descrizione: testo tesi
Dimensione
40.72 MB
Formato
Adobe PDF
|
40.72 MB | Adobe PDF | Visualizza/Apri |
|
2026_10_Pachioli_Executive_Summary.pdf
accessibile in internet per tutti a partire dal 01/10/2029
Descrizione: testo executive summary
Dimensione
2.27 MB
Formato
Adobe PDF
|
2.27 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/10589/267055