CL_IV_L: Clinical Medicine IV. Lectures
Bettina Katalin Budai1, Ambrus Tóth1, Petra Borsos1, Veronica Grace Frank1, Sonaz Shariati1, Bence Fejér1, Anikó Folhoffer2, Ferenc Szalay2, Viktor Bérczi1, Pál Maurovich Horvat1, Pál Novák Kaposi1
1 Department of Radiology, Medical Imaging Centre, Faculty of Medicine, Semmelweis University, Budapest
2 Department of Internal Medicine and Oncology, Faculty of Medicine, Semmelweis University, Budapest
CT texture analysis (CTTA) has been used successfully for the quantitative assessment of tissue heterogeneity in both benign and malignant lesions of various organs.
Our study aimed to develop CTTA-based machine learning models, which can be used for the non-invasive evaluation of liver fibrosis.
In this study, portal venous phase CT scans of 32 chronic hepatitis patients with liver fibrosis was retrospectively collected. The scans were performed on either a 16- or a 64-slice CT scanner. The fibrosis stages were determined based on shear wave elastography. After the manual delineation of the anatomic liver segments, three-dimensional CTTA was performed on each segment with a total of 1117 radiomic parameters (RPs). Pearson’s correlation analysis was used to filter out the highly-correlated RPs (r>0.95). The remaining 453 RPs were used for k-means clustering and in hierarchical cluster analysis (HCA). The best performing RPs were selected with recursive feature elimination. For the machine learning-based classification, the segments were split between the training and test sets in equal proportion (analysis I) or based on the scanner type (analysis II). The continuous variables were compared with ANOVA and post hoc Tukey’s tests. The performance of the models was evaluated based on the area under the receiver operating characteristic curve (AUC).
K-means clustering and HCA divided the segments into four main clusters. Cluster-4 showed a significantly higher average CT density (110.0HU ± 10.1HU) compared to the other clusters (c1: 96.1HU ± 11.3HU, p<0.0001; c2: 90.8HU ± 16.8HU, p<0.0001; c3: 93.1HU ± 17.5HU, p<0.0001); but there was no difference in the distribution of liver stiffness or scanner types. The trained random forest classifier showed an excellent prediction performance in differentiating between low-grade and high-grade fibrosis in both analysis I and II (AUC = 0.90 vs. AUC = 0.88). Meanwhile, the support vector machine classifier achieved an excellent performance in analysis II (AUC = 0.91) and an acceptable prediction rate in analysis I (AUC = 0.76).
In conclusion, CTTA could be used successfully to differentiate high-grade from low-grade fibrosis irrespective of the imaging platform.
PNK was supported by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences (386/2017).
Semmelweis University, Károly Rácz Doctoral School of Clinical Medicine