Automated Detection of Liver Steatosis in Ultrasound Images Using Convolutional Neural Networks

Artificial Intelligence for Detecting and Quantifying Fatty Liver in Ultrasound Images: A Systematic Review

“…Typically, ROIs are chosen manually by an expert or automatically using methods like a self-organizing map [ 53 ]. In this research, some studies used the complete image as the ROI with a size of 434 × 636 pixels [ 60 , 61 , 63 , 64 ], 800 × 600 pixels [ 71 ], 960 × 720 pixels [ 51 ], or 1024 × 1024 pixels [ 39 ]. Other studies selected larger ROIs with 500 × 500 pixels [ 36 , 37 , 53 ].…”

“…Features such as maximum probability [ 38 , 52 , 53 , 59 , 63 , 78 , 80 ], uniformity [ 38 , 52 , 53 , 59 , 60 , 63 , 70 , 80 ], entropy [ 37 , 38 , 49 , 52 , 59 , 63 , 66 , 70 , 71 , 73 , 74 , 76 , 80 , 82 ], contrast [ 38 , 49 , 52 , 59 , 63 , 65 , 66 , 70 , 71 , 73 , 74 , 76 , 78 , 80 ], run-length uniformity [ 38 , 52 , 70 , 80 ], attenuation [ 43 , 73 , ...…”

“…There are many methods used to extract features from US images. Among the extraction methods frequently used in studies with higher performance are the grey-level co-occurrence matrix method [ 38 , 52 , 54 , 58 , 63 , 71 , 76 , 79 , 80 , 83 ], grey-level difference statistics [ 48 , 49 , 66 ], grey-level run-length matrix [ 38 , 48 , 49 , 58 ], spatial grey-level dependence [ 47 , 48 , 49 , 53 , 65 , 66 ], wavelet packet transform [ 55 , 56 , 57 ], and first-order statistics [ 52 , 58 , 79 ].…”

“…Even though not all studies provided information about the methods used to manage data, most studies used cross-validation. Twelve studies implemented 10-fold cross-validation on the testing and training datasets [ 36 , 38 , 39 , 42 , 43 , 52 , 58 , 60 , 62 , 67 , 68 , 70 ], four studies used five-fold cross-validation [ 59 , 63 , 66 , 69 ], one study used four-fold cross-validation [ 77 ], two studies used three-fold cross-validation [ 41 , 54 ], one study used two-fold cross-validation [ 40 ], and one study used one-fold cross-validation [ 74 ]. In addition, four studies mentioned that a leave-one-out cross-validation (LOOCV) method was used [ 55 , 56 , 57 , 79 ].…”

“…Two studies used Inception-v3 as a classification model [ 51 , 84 ], and two studies used VGG-16 to classify US images [ 50 , 69 ]. Five studies used a residual neural network (ResNet) with the following version variations: Inception-ResNet-v2 [ 63 ], multi-scale two-dimensional mid-fusion ResNet [ 62 ], ResNet-18 [ 45 , 51 ], and ResNet-50 v2 [ 81 ]. Finally, in one study, a SVM was used to classify images after a CNN was used to extract features [ 59 ].…”