Subpixel segmentation of borehole fractures from low resolution Doppler ultrasound images using machine learning

Abstract

Ultrasonic imaging is utilized in oil and gas, and geothermal boreholes using various logging tools. The imaging resolution of these tools is physically limited by the point spread function (PSF) of the imaging system. It is of interest to be able to accurately resolve features that are blurred out due to the PSF limit. This work evaluates a convolutional neural network (CNN), adapted from medical ultrasound imaging, to obtain subpixel estimates of fracture shape and area in boreholes from ultrasonic power Doppler images. The CNN was trained using a dataset generated using simplified ultrasonic simulations, ignoring effects due to fluid dynamics. The performance of the CNN model was tested on experimental scans obtained using a laboratory setup using a water based drilling mud. The CNN was able to estimate the fracture areas with a significantly lower mean absolute error of 5.0±3.9mm2, compared to 22.9±1.7mm2 using the conventional thresholding method. The CNN also enabled the estimation of fracture shape and area using a single frame of the power Doppler image, compared to about 30 frames required for the thresholding method, facilitating a possible increase in scanning speed of the logging tool. Although, the CNN in this work is applied on power Doppler images from borehole fractures, the model can be easily adapted for similar other applications, such as ultrasonic pulse-echo imaging of boreholes, cement bond evaluation, perforation evaluation in production logging, etc.