Automatic Prostate Volume estimation on Transabdominal Ultrasound using Deep Neural Networks

Abstract

Prostate cancer is the most commonly diagnosed cancer in the Netherlands. Accurate assessment of the prostate volume (PV) is a crucial step in prostate cancer (PCa) screening and risk-stratification. In standard clinical care, the PV is obtained by measuring the prostate dimensions with the aid of transrectal ultrasound (TRUS). However, rectal examination is characterised with patient discomfort, for which the feasibility of transabdominal ultrasound (TAUS) is explored, as a more accessible and patient-friendly alternative. However, manual PV measurements are prone to inter-observer variability and require operator training. This study aims to improve the accessibility, complexity, and robustness of PV measurements by developing a framework to automatically estimate the PV based on TAUS acquisitions. The primary components of the framework comprise two deep neural networks, developed for prostate segmentation on axial and sagittal TAUS images, and an algorithm that extracts the prostate's diameters on which the PV is calculated.

During this study, a new prostate dataset is developed, comprising sagittal and transverse TAUS image acquisitions of 100 participants, and reference PV measurements based on TRUS and MRI are collected. First, the feasibility of TAUS for manual PV estimation is explored, and the inter-method agreement between TAUS, TRUS and MRI is analysed in Bland Altman diagrams. Additionally, all TAUS acquisitions are assessed on image quality.
Secondly, three deep neural networks (using the nnU-Net framework) are developed to segment the prostate on sagittal and/or axial TAUS images. All models, are trained and validated on TAUS image data of 52 participants. Additionally, an algorithm is designed to predict the prostate diameters when prostate segmentations serve as input. To this extend, the PV is estimated according to the widely used Ellipsoid formula. The proposed algorithm is evaluated on input ground-truth segmentations of 42 participants. Essentially, the segmentation models combined with the proposed algorithm result in a framework to automatically estimate the PV on TAUS. Finally, it is tested on unseen TAUS acquisitions of 17 participants, whereby the predicted PV is compared to reference PV measurements on MRI.

Our results show an average volume difference of 3.0 ± 17.6 ml when manual PV estimation on TAUS is compared to MRI. When manual PV estimation on TRUS is compared to MRI, an average volume difference of 12.3 ± 18.8 ml is obtained. The developed segmentations models segment the middle region of the prostate on TAUS with an average DSC = 0.91 ± 0.06 and DSC = 0.83 ± 0.09 for axial and sagittal TAUS images respectively. When the entire prostate region was evaluated, a lower model performance was observed, whereby the prostate was segmented with a DSC of 0.76 ± 0.09 in the axial imaging-plane and DSC of 0.68 ± 0.21 in the sagittal imaging-plane. The algorithm for automatic diameter extraction showed good correspondence with manually assigned prostate diameters on TAUS.
When the segmentation models and the algorithm are utilised for automatic PV estimation, an average volume difference of 2.5 ± 10.2 ml was observed, compared to MRI reference volumes. Ultimately the PV was predicted with a volume difference < 25 compared to MRI in 14 out of 17 test cases. 

The results of this study show that it is possible to obtain PV measurements using TAUS that are comparable to those obtained with MRI. Moreover, the variability related to PV measurements using TAUS seem unrelated to TAUS image quality, indicating that manual PV measurements can be performed, even when unfavorable patient characteristics limit the image quality. Still, it is important to note that proper operator training for TAUS examination is essential. The proposed framework for automatic PV estimation on TAUS acquisitions shows good correspondence with MRI reference volumes. Thus expanding the possibilities of PCa risk-stratification, whereby robust, and straightforward PV estimations are desired. In order to adopt the framework for standard clinical care, further research is required on a larger cohort to investigate the generalizability of the framework and ensure reliable results on all future patients.