TY - JOUR
T1 - Automated Detection of Liver Steatosis in Ultrasound Images Using Convolutional Neural Networks
AU - Mohammad, Umar Farooq
AU - Almekkawy, Mohamed
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - The most often used imaging modality for the diagnosis of fatty liver disease is ultrasound imaging. If there is more than 5% fatty hepatorenal steatosis, it is deemed malignant. The traditional procedures for classifying hepatic steatosis mainly need the use of trained radiologists. Methods such as the Hepatorenal Index (HI) and Gray-level co-occurrence matrix (GLCM) are routinely employed. A Convolutional Neural Network (CNN)-based strategy to classifying malignant and benign fatty livers from ultrasound images is presented in this research. The pre-trained Inception Resnet is utilized for transfer learning on B-mode ultrasound liver pictures for categorization. It was first trained on the ImageNet dataset. The softmax activation function is employed as an output layer of the fully connected network to calculate the class-wise probabilities, and the features recovered by the Inception Resnet are provided as input to the fully connected network. We used a 550-image open-source ultrasound liver dataset comprising 170 normal and 380 malignant samples. Our tests revealed that the Inception Resnet has a classification accuracy of 98.48%, whereas the HI and the GLCM methods have accuracies of 90.9 and 85.4 percent, respectively.
AB - The most often used imaging modality for the diagnosis of fatty liver disease is ultrasound imaging. If there is more than 5% fatty hepatorenal steatosis, it is deemed malignant. The traditional procedures for classifying hepatic steatosis mainly need the use of trained radiologists. Methods such as the Hepatorenal Index (HI) and Gray-level co-occurrence matrix (GLCM) are routinely employed. A Convolutional Neural Network (CNN)-based strategy to classifying malignant and benign fatty livers from ultrasound images is presented in this research. The pre-trained Inception Resnet is utilized for transfer learning on B-mode ultrasound liver pictures for categorization. It was first trained on the ImageNet dataset. The softmax activation function is employed as an output layer of the fully connected network to calculate the class-wise probabilities, and the features recovered by the Inception Resnet are provided as input to the fully connected network. We used a 550-image open-source ultrasound liver dataset comprising 170 normal and 380 malignant samples. Our tests revealed that the Inception Resnet has a classification accuracy of 98.48%, whereas the HI and the GLCM methods have accuracies of 90.9 and 85.4 percent, respectively.
UR - http://www.scopus.com/inward/record.url?scp=85122873789&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85122873789&partnerID=8YFLogxK
U2 - 10.1109/IUS52206.2021.9593420
DO - 10.1109/IUS52206.2021.9593420
M3 - Conference article
AN - SCOPUS:85122873789
SN - 1948-5719
JO - IEEE International Ultrasonics Symposium, IUS
JF - IEEE International Ultrasonics Symposium, IUS
T2 - 2021 IEEE International Ultrasonics Symposium, IUS 2021
Y2 - 11 September 2011 through 16 September 2011
ER -