Automated liver tissue segmentation in point of care ultrasound b-mode images using U-Net

December 15, 2021


Introduction/Purpose: We have recently developed and demonstrated that tissue acoustic properties can facilitate early detection of liver cancer during period abdominal ultrasound surveillance exams in high risk patients. Automated segmentation of the liver is a necessary step in increasing the accuracy, sensitivity, and specificity of Machine Learning Models which detect cancer using the lower quality data generated from Point-of-Care Ultrasound (PoCUS). We have developed deep learning algorithms which can successfully segment the data generated by these PoCUS systems. Methods: A total of 1000 patients from 4 collaboration sites were recruited, though for this study only 25 of which were manually segmented were used.. For this HIPPA compliant study, all data was obtained through IRBs approved at the local data acquisition site; all patients were consented before data acquisition. The patient's mean age and BMI were 43(11 STD) and 23 (5 STD). All patients were scanned using the same acquisition protocol, capturing the same 5 scan visualization windows of the liver using a PoCUS system, with capabilities to capture raw signal data. We use a deep learning algorithm to automatically segment out the liver. The initial version of the U-net Deep learning architecture was trained on a subset of our data containing B-mode images from 20 patients (500 frames) and tested on a dataset with 5 patients (125 frames). The Regions of interest that served as training and testing labels for the algorithm were outlined on an image-by-image basis by expert sonographers. U-Net is a fully convolutional neural network that consists of a block of contraction layers followed by a block of expansion layers. The architecture contains feature channels to relay context information to higher resolution layers. U-Net has been used at multiple medical segmentations tasks in which it performs particularly well even with small datasets. Results: The architecture successfully segmented the images both in the train and test datasets with a great degree of accuracy and generalization. Test accuracies surpassed 86% with a sensitivity of 80% and specificity of 95% and results can be further improved by applying thresholds to the algorithm’s output. Conclusion: This automated segmentation algorithm is a clear next step in the improvement of any automated disease detection method in the liver. Once in place it will greatly aid in the early screening of all liver diseases and improve the outcomes for at risk patients.

Read the full paper here.