Deep Learning-Based Automatic Measurement of Prostate Volume Using Two- Dimensional Transrectal Ultrasound: A Pilot Study

Abstract

Abstract Purpose The aim of this study is to automatically measure the prostate volume based on two-dimensional transrectal ultrasound (TRUS) images using deep learning. Methods A total of 1,645 ultrasound images were collected from 110 patients and partitioned into training and testing datasets at a 10:1 ratio. We introduced a method for automated prostate volume measurement using deep learning techniques on 2-D TRUS images. Initially, we applied a VGG-19-based model to classify the cross-sectional TRUS images into two types – transverse and sagittal images. These two groups of the categorized images were respectively fed into a U-Net-based model for segmentation. From the U-Net-based model, we obtained segmented images that are used to measure the prostate’s length and volume via an ellipsoid method. The measured volumes from our model were quantitatively evaluated by radiologist-measured results; the classification network was assessed based on accuracy and the segmentation network was assessed using intersection over union metrics, respectively. Results The classification network showed an accuracy of 99.35%, and the segmentation network exhibited a mean intersection over union value of 90.88%. The average error rate between the measured volume of the proposed method and the volume measured by the clinical assessment is 9.17%. Conclusions We have demonstrated that 2-D TRUS images, obtained through routine clinical diagnosis of prostate, can be effective and reliable in measurement of prostate volumes through a deep-learning-based approach with the ellipsoid formula. Our method can potentially address the previous challenges related to patient-dependent prostate shapes, ambiguous brightness patterns in TRUS, and inconsistencies among different operators when performing manual measurements.