Prostate cancer diagnosis typically combines MRI for lesion detection with ultrasound (US) guidance for targeted biopsy. Although 3D US acquisitions are often performed be- fore or during the intervention, they are not routinely exploited for continuous intraopera- tive navigation. In practice, the physician must mentally align real-time 2D slices with the reference 3D anatomy, which can be error-prone and highly dependent on experience. To mitigate this limitation, automatic 2D/3D registration methods have been explored, but challenges such as probe motion complexity and anatomical symmetry remain unsolved. In this work, we propose a multitask deep learning model designed to address these issues by jointly learning prostate segmentation as a spatial context and transformation regres- sion for probe tracking. The architecture employs a ResNet-34 shared encoder connected to two decoders: a U-Net branch that outputs segmentation masks, and a multilayer per- ceptron (MLP) branch that estimates transformation parameters. Batch normalization layers were systematically replaced by instance normalization layers to prevent the model from losing its learned referential due to unevenly oriented biopsy volumes. Temporal con- sistency is enforced by reinjecting previous predictions into the model input at each time step. Experiments conducted on 17-second synthetic trajectories show that, without drift correction, the best configuration achieves a mean target registration error of 7.47 ± 6.94 mm. Moreover, 87% of trajectories of 350 frames remain within a 7 mm error threshold. These findings suggest that segmentation-informed context can provide a robust spatial reference to stabilize registration and solve the symmetry ambiguity, while the single- model design offers a partial solution to symmetry-related ambiguities. Despite residual errors on long sequences, this study demonstrates the potential of multitask learning for sensorless probe tracking and real-time guidance in prostate biopsy procedures.
La diagnosi del cancro alla prostata combina tipicamente la risonanza magnetica per l’individuazione delle lesioni con la guida ecografica (US) per la biopsia mirata. Sebbene le acquisizioni ecografiche 3D vengano spesso eseguite prima o durante l’intervento, non vengono sfruttate sistematicamente per la navigazione intraoperatoria continua . In prat- ica, il medico deve allineare mentalmente le sezioni 2D in tempo reale con l’ anatomia 3D di riferimento, il che può essere soggetto a errori e dipendere fortemente dall’esperienza. Per mitigare questa limitazione, sono stati studiati metodi di registrazione automatica 2D/3D, ma sfide quali la complessità del movimento della sonda e la simmetria anatomica riman- gono irrisolte. In questo lavoro, proponiamo un modello di deep learning multitasking progettato per affrontare questi problemi apprendendo congiuntamente la segmentazione della prostata come contesto spaziale e la regressione della trasformazione per il traccia- mento della sonda. L’architettura impiega un codificatore condiviso ResNet-34 collegato a due decodificatori: un ramo U-Net che produce maschere di segmentazione e un ramo multilayer perceptron (MLP) che stima i parametri di trasformazione. I livelli di Batch normalization sono stati sistematicamente sostituiti da livelli di Instance Normalisation per impedire al modello di perdere il riferimento appreso a causa di volumi di biopsia orientati in modo non uniforme. La coerenza temporale viene garantita reiniettando le previsioni precedenti nell’input del modello ad ogni passo temporale. Gli esperimenti con- dotti su traiettorie sintetiche di 17 secondi dimostrano che, senza correzione della deriva, la configurazione migliore raggiunge un errore medio di registrazione del target pari a 7,47 ± 6,94 mm. Inoltre, l’87% delle traiettorie di 350 fotogrammi rimane entro una soglia di errore di 7 mm. Questi risultati suggeriscono che il contesto informato dalla segmentazione può fornire un solido riferimento spaziale per stabilizzare la registrazione e risolvere l’ambiguità di simmetria, mentre il design a modello singolo offre una soluzione parziale alle ambiguità legate alla simmetria. Nonostante gli errori residui su sequenze lunghe, questo studio dimostra il potenziale dell’apprendimento multitasking per il trac- ciamento senza sensori e la guida in tempo reale nelle procedure di biopsia prostatica.
Deep learning-assisted ultrasound navigation for prostate biopsies
ROLLAND, TANGUY JACQUES MARIE
2024/2025
Abstract
Prostate cancer diagnosis typically combines MRI for lesion detection with ultrasound (US) guidance for targeted biopsy. Although 3D US acquisitions are often performed be- fore or during the intervention, they are not routinely exploited for continuous intraopera- tive navigation. In practice, the physician must mentally align real-time 2D slices with the reference 3D anatomy, which can be error-prone and highly dependent on experience. To mitigate this limitation, automatic 2D/3D registration methods have been explored, but challenges such as probe motion complexity and anatomical symmetry remain unsolved. In this work, we propose a multitask deep learning model designed to address these issues by jointly learning prostate segmentation as a spatial context and transformation regres- sion for probe tracking. The architecture employs a ResNet-34 shared encoder connected to two decoders: a U-Net branch that outputs segmentation masks, and a multilayer per- ceptron (MLP) branch that estimates transformation parameters. Batch normalization layers were systematically replaced by instance normalization layers to prevent the model from losing its learned referential due to unevenly oriented biopsy volumes. Temporal con- sistency is enforced by reinjecting previous predictions into the model input at each time step. Experiments conducted on 17-second synthetic trajectories show that, without drift correction, the best configuration achieves a mean target registration error of 7.47 ± 6.94 mm. Moreover, 87% of trajectories of 350 frames remain within a 7 mm error threshold. These findings suggest that segmentation-informed context can provide a robust spatial reference to stabilize registration and solve the symmetry ambiguity, while the single- model design offers a partial solution to symmetry-related ambiguities. Despite residual errors on long sequences, this study demonstrates the potential of multitask learning for sensorless probe tracking and real-time guidance in prostate biopsy procedures.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_Report_Tanguy_Rolland_2025_V2.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
6.85 MB
Formato
Adobe PDF
|
6.85 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/10589/243953