Keywords:
Cirrhosis, Segmentation, Computer Applications-Detection, diagnosis, CT, Liver, Artificial Intelligence, Abdomen
Authors:
S. L. Mihalcioiu1, R. Remtulla2, O. ciga2, C. H. MO3, M. D. A. Attarian3, P. Savadjiev3, S. BHATNAGAR3, C. Reinhold1, J. J. R. Chong1; 1Montreal, QC/CA, 2Montreal, Quebec/CA, 3Montreal/CA
DOI:
10.26044/ecr2019/C-3419
Methods and materials
Study Design:
Patients were selected and a database was created of unenhanced and Portal Venous CT images through selection of mild,
moderate and severe cases of hepatic steatosis as defined by the unenhanced phase appearance.
Normal control cases of normal density healthy livers were used for each of the unenhanced CT and portal venous CT groups.
The unenhanced CT phase images were used as an internal reference standard for hepatic steatosis.
1.5 cm2 Regions of Interest (ROI) were taken from hepatic segments V,
VI,
VII,
and VIII from the unenhanced and portal venous phases.
To measure splenic attenuation three 1.5 cm2 ROIs were taken from the superior,
middle,
and inferior segments of the spleen.
Images selected in the Portal Venous and Unenhanced phase were used in the training dataset.
A pre-trained transfer-learning ‘V-NET’ 3D convolutional neural network was utilized to segment the whole liver organ boundary masking the image to just liver parenchyma.
This masked 3D volume served as the the training and validation image dataset for training a final 3D CNN classifier to predict for normal/mild/moderate/severe steatosis.
The final algorithm was validated on a reserved set of test examinations,
composed of entirely independent patients from the training and validation sets.
Standard statistical methods were used to generate the classification accuracy and receiver operating characteristic (ROC) curve of the final trained network with the Unenhanced CT image appearance used as the reference standard.
The CNN was evaluated under two use cases unknown unenhanced and unknown portal venous CT abdomen and pelvis volumes to determine if the algorithm could successfully segment the liver parenchyma,
and distinguish between grades of A-none,
B-mild,
C-moderate,
and D-severe steatosis.
The optimal ROC sensitivity and specificity of the CNN algorithm’s hepatic steatosis detection were calculated and compared with similar prior studies on human inter-rater variability on hepatic steatosis.
A polynomial regression model was fit on the 40 Training cases to predict Average Plain Density from Average PVP Density.
The overall training and validation methodology is depicted in Fig. 1 and Fig. 2.
Inclusion and Exclusion Criteria:
Inclusion: All axial abdominal CT scans of the Portal Venous phase as well as the Unenhanced phase from patients with hepatic steatosis and normal livers as a control dating from 2006-2018.
Exclusion: All examinations with confounding dense artifact,
insufficient technical quality or confounding liver/spleen pathology were excluded from the training and validation sets.
These comprised hemosiderosis,
focal solid liver lesions,
liver mets or a prior history of malignancy.
Small incidental likely benign lesions less than 2 cm were permitted,
but larger lesions that significantly alter the density of the liver organ as a whole were excluded.
Sample Size:
80 multiphasic single-energy CT examinations were selected,
an even balance of mild/moderate/severe steatosis cases,
and 20 normal controls.
All volumes underwent standardized multi-ROI evaluation to determine an average Plain and PVP liver density,
and were split into 40 Training and 40 Test cases.