Automatic Prostate Volume Estimation in Transabdominal Ultrasound Images

TL;DR

The paper presents a deep-learning framework for automatic prostate volume estimation from transabdominal ultrasound TAUS videos. It deploys two plane-specific nnU-Net segmentation models to segment the prostate in axial and sagittal TAUS frames, extracts three diameters from the segmentations, and computes PV via the ellipsoid formula $PV_{Ellipsoid} = D_{frontal} \times D_{longitudinal} \times D_{sagittal} \times \pi / 6$. Using MRI-derived PV as a reference, the approach achieves mean volumetric errors around $-5$ to $0$ mL with relative errors typically in the $5$–$15\%$ range, though sagittal-plane segmentation is more challenging. The study demonstrates promising non-invasive PV estimation from TAUS, highlighting strong axial-plane performance and potential for broader clinical adoption with improved data and augmentation. Overall, the framework offers a feasible path toward non-invasive PV-based risk stratification for early prostate cancer detection.

Abstract

Prostate cancer is a leading health concern among men, requiring accurate and accessible methods for early detection and risk stratification. Prostate volume (PV) is a key parameter in multivariate risk stratification for early prostate cancer detection, commonly estimated using transrectal ultrasound (TRUS). While TRUS provides precise prostate volume measurements, its invasive nature often compromises patient comfort. Transabdominal ultrasound (TAUS) provides a non-invasive alternative but faces challenges such as lower image quality, complex interpretation, and reliance on operator expertise. This study introduces a new deep-learning-based framework for automatic PV estimation using TAUS, emphasizing its potential to enable accurate and non-invasive prostate cancer risk stratification. A dataset of TAUS videos from 100 individual patients was curated, with manually delineated prostate boundaries and calculated diameters by an expert clinician as ground truth. The introduced framework integrates deep-learning models for prostate segmentation in both axial and sagittal planes, automatic prostate diameter estimation, and PV calculation. Segmentation performance was evaluated using Dice correlation coefficient (%) and Hausdorff distance (mm). Framework's volume estimation capabilities were evaluated on volumetric error (mL). The framework demonstrates that it can estimate PV from TAUS videos with a mean volumetric error of -5.5 mL, which results in an average relative error between 5 and 15%. The introduced framework for automatic PV estimation from TAUS images, utilizing deep learning models for prostate segmentation, shows promising results. It effectively segments the prostate and estimates its volume, offering potential for reliable, non-invasive risk stratification for early prostate detection.

Figures (5)

• Figure 1: Overview of the study. A dataset of 100 patients has been acquired between August 2023 and February 2024. Prostates of all patients were scanned with a TAUS transducer, and at least one US video from the axial and one along the sagittal imaging plane was acquired. As a result of a quality check (improper transducer handling or technical issues with the scanner), US videos from 38 patients have been excluded. US videos from the remaining patients were split into training and testing sets. Two deep-learning models, one per imaging plane, were trained to automatically segment the prostate from TAUS frames. Prostate diameters were extracted and used to calculate the prostate volume using the ellipsoid formula and then compared to their respective reference-standard values (MRI).
• Figure 2: Example of manual delineation process. For each frame in the US videos, a clinician provided a boundary annotation of the prostate. At the same time, prostate diameters were measured in the TAUS videos. A) Ground truth segmentation mask based on boundary annotation. B) Extracted diameters, in this example a longitudinal and a frontal diameter. C) Distributions of the manually delineated diameters. Longitudinal and frontal diameters are extracted from the axial US videos and the sagittal diameter from sagittal ones.
• Figure 3: Model performance summarized in boxplots for each fold in the cross-validation of $Model_{Ax}$ and $Model_{Sag}$ on the $TrainSet_{Ax}$ and $TrainSet_{Sag}$ splits, respectively. For each fold, Dice, Dice Mid-Plane, and HD Mid-Plane are reported.
• Figure 4: Qualitative analysis of the results from the models for automatic prostate segmentation. Segmentation results from US videos along both imaging planes from 3 patients. For each axial sweep, frames from the base, mid-plane, and apex are shown and for each sagittal sweep, frames from left, center, and right are shown. Both models achieved good segmentation performances on the mid-plane frames. $Model_{Ax}$ shows improved performances over $Model_{Sag}$, especially when segmenting from US frames of the prostate extremities.
• Figure 5: A: Bland Altman plot comparing the predicted prostate volume (PV) measurements and reference-standard volume measurements based on MRI, on 10 patients in the Test set (\ref{['sec:partitioning']}). B: Distribution of the volume estimation error on the test set (\ref{['sec:partitioning']}) relative to prostate dimension computed as average between reference-standard and estimated.