Uncertainty-aware hybrid model for reliable navigation guidance in prostate biopsy

Prostate biopsy - performed under ultrasound (US) guidance - is the only confirmatory procedure for diagnosing prostate cancer. If several systems aiming at improving the accuracy of tissue sampling for histopathological analysis exist, they often fail to adequately account for patient / tissue motion and deformation during the procedure or are not suitable for real-time guidance towards a pre-defined target.

This work proposes a real-time assistance method to help urologists navigate between biopsy sites leveraging both 2D US and 3D US information.

The proposed hybrid framework combines three ResNet-18-based sub-networks with a Kalman filter to track the pose of the current 2D US frame within a reference 3D US volume acquired during the intervention. Two of the networks predict relative rigid transformations from pairs of US frames or frame/slice combinations extracted from the 3D volume. A third network predicts the registration error from a frame/slice pair and acts as a large-error detector. This error prediction module is integrated directly into the iterative registration process, to ensure robust tracking in the presence of tissue deformation or image artifacts, while also providing an uncertainty estimate of registration quality at each time step. The proposed framework was trained and evaluated on complex biopsy trajectories generated from a dataset of 847 3D US volumes acquired during real transrectal prostate biopsy procedures from 97 patients. The dataset was divided into training (78 patients, 669 volumes), validation (10 patients, 115 volumes), and test (9 patients, 63 volumes) sets. Performance was compared with an image-correlation-based baseline method using clinically relevant evaluation metrics, including the mean surface registration error (mean SRE) over 25-s trajectories, as well as the percentage of registered frames with an SRE below 2.5 mm and 5 mm. Statistical significance was assessed using Wilcoxon signed-rank tests (adjusted with Holm-Bonferroni correction) and 95% confidence bootstrap intervals (CI 95%) across the test cohort. Effect sizes were quantified with rank-biserial correlation (RBC).

Compared with the baseline method, the proposed approach achieved a significant reduction in the mean SRE over 25-s trajectories, decreasing from $3.1[2.2,4.9]$ to $2.2[1.8,2.5]$  mm (median [CI 95%], p = 0.016, RBC = 0.96). In addition, the proposed large-error detector - used for uncertainty estimation - achieved a balanced accuracy score of 0.81 for classifying registration errors above or below a given threshold of 5 mm.

These results demonstrate the potential of the proposed framework to improve the accuracy and robustness of real-time prostate biopsy navigation guidance.