Comparison of histopathological liver fibrosis stage in patients diagnosed with chronic hepatitis B and C using radiomics analysis

In this study, we aimed to evaluate liver fibrosis and necroinflammation in chronic hepatitis B and C patients using radiomics features, thus protecting patients from biopsy complications and providing clinicians with an easier, more reproducible, non-invasive diagnostic method. After obtaining approval from the local ethics committee for the study, patients who underwent liver biopsy at our hospital between January 2015 and January 2024 and were histopathologically diagnosed with chronic hepatitis B and C were retrospectively evaluated. Patients who had a contrast-enhanced abdominal magnetic resonance (MRI) scan within 6 months and laboratory tests (albumin, ALT, AST, platelet count, INR, PT) within 1 month before the biopsy were included in the study. After applying the inclusion and exclusion criteria, 50 patients with different stages of liver fibrosis were included in the study. Additionally, 30 patients without liver disease or liver lesions were included in the study with a fibrosis score of 0. Thus, a total of 80 patients with liver fibrosis scores ranging from 0 to 6 were included in the study. Axial T2-weighted (T2W), portal, and late phase contrast-enhanced T1-weighted (T1W) sequences from these patients' contrast-enhanced abdominal MRI scans were loaded into the 3D Slicer program. The liver was segmented 3D volumetrically using this program. The PyRadiomics 3.0.1 software package was then used to extract radiomics features from the segmented VOI (Volume of Interest). 130 radiomics features were extracted for each sequence. Features that did not contain a single numerical value or that text were excluded. 106 radiomics features were used for each sequence. A total of 318 radiomics features were obtained from measurements performed on three different sequences. After extracting radiomics features, before proceeding to the model creation phase, patients were randomly assigned to training and test groups using a 3:1 (60:20) distribution method. Model creation and training were performed in the training group, and the performance evaluation of the created model was performed in the test group. Using the Random Forest algorithm, the top seven variables providing the most prominent information contribution were selected from the radiomics features. A radiomics model was created using the selected seven variables. A clinical model was created using laboratory data, and a combined model was created by gathering these data. To evaluate the success of radiomics features in predicting fibrosis, we divided the patients into binary groups based on fibrosis stage and HAI. We created five different groups: Group 1: Fibrosis 0-2 vs. 3-6, Group 2: Fibrosis 0 vs. 1-6, Group 3: Fibrosis 0-1 vs. 2-6, Group 4: HAI 0-12 vs. 13-18, and Group 5: HAI 0-5 vs. 6-18. For each group, a model was created using six different machine learning methods to evaluate the ability to distinguish between the groups. The machine learning algorithms used were Logistic regression, Random Forest, SVM (Linear), SVM (RBF), KNN, Naive Bayes, and XGBoost. In the performance evaluation process, criteria such as sensitivity, precision, accuracy, F1 score and area under the ROC curve (AUC) were used. The most successful results in distinguishing fibrosis stages 0–2 and 3–6 were achieved in clinical models.Particularly,the SVM (RBF) algorithm stands out with 70% accuracy, an AUC of 0.667, and an F1-score of 0.5. While radiomics models demonstrated similar accuracy and AUC, many models had F1-score values below 0.400. All models demonstrated high success in distinguishing fibrosis stages 0 and 1–6, with radiomics and combined models demonstrating excellent performance. Random Forest and Naive Bayes algorithms achieved full performance across all metrics with 100% accuracy, an AUC of 1.000, and an F1-score of 1.000. All models demonstrated high performance in distinguishing fibrosis stages 0–1 and 2–6. However, radiomics models performed significantly better. The SVM (RBF) model achieved the best results with 98% accuracy, an AUC of 0.98, and an F1-score of 0.952. Among the radiomics models, the SVM (Linear) algorithm was the most successful model in distinguishing HAI 0–12 and 13–18 with 95% accuracy, 0.961 AUC and 0.857 F1-score. Among the models created for the discrimination of HAI 0–5 and 6–18, Random Forest, one of the combined models, showed the highest performance with 95% accuracy, 0.98 AUC and 0.952 F1-score. In conclusion, radiomics models demonstrated high success in identifying patients with and without fibrosis (F0 vs. F1-6) and ≥F2. They also classified cases with HAI ≥13 and HAI ≥6 with high accuracy. The 2023 Turkish Hepatitis B Diagnostic and Therapeutic Guidelines specify F ≥2 and HAI ≥6 as treatment indications. Our model allows non-invasive identification of this patient group. However, it failed to demonstrate sufficient accuracy in distinguishing between F0-2 and F3-6, and the clinical model was found to be more successful. Nevertheless, radiomics models have the potential to provide insight into the level of fibrosis and necroinflammation in routine patient follow-up. However, prospective multicenter studies are needed. In our study, we observed significant differences in the performance of the six different machine learning algorithms we used. This demonstrates the importance of choosing the right algorithm in radiomics studies. Among the algorithms we used, Random Forest (RF) and Support Vector Machine (SVM) were generally more successful. Keywords: Hepatitis, Liver Fibrosis, HAI, Radiomics, Machine Learning, Segmentation