Prediction of metastatic LYMPH nodes in malignant melanoma patients by texture analysis and machine learning performed in F18-FDG pet and computerized tomography

AIM: Since the early 1990s, lymphatic mapping and sentinel lymph node biopsy (SLNB) in cutaneous malignant melanoma patients with AJCC staging T1b, T2, T3 and clinically lymph node negative; is routinely used as a staging procedure1. However, latest imaging techniques such as computed tomography texture analysis are thought to be useful in predicting lymph node metastasis and prognosis of patients in oncologic imaging. In this study, texture analysis parameters of lymph nodes in computerized tomography and Positron Emission Tomography images and metabolic characters in PET-CT will be correlated with the pathology of postoperative lymph nodes in malignant melanoma patients. In this study, we aimed to detect metastatic lymph nodes in patients with malignant melanoma using texture analysis and machine learning. MATERIALS AND METHODS: Volumetric segmentation was performed by two radiologists by placing a free-hand region of interest (ROI) on Canon's 'OLEA' software on PET-CT images. We analysed the significant parameters of texture analysis in metastatic lymph nodes. During the feature selection phase, reproducible features were selected with intra-class correlations (ICC) analysis (threshold 0.75). A new combined group was created by combining reproducible features from PET and CT images. According to the Pearson correlation coefficients, the features that did not correlate more than 0.7 with each other were included in the third stage. At the last stage, the features were selected using the wrapper-based recursive feature selection algorithm. Models were created using Support vector machine (SVM) and Logistic Regression (LR) algorithms. RESULTS: A total of 44 malignant melanoma patients underwent diagnostic 18F-FDG PET/CT imaging to evaluate lymph node metastasis to determine the stage of the disease. After SLNB and/or lymph node dissection 24 of them had been reported with regional lymph node metastasis. In texture analysis 111 features of the lymph nodes were extracted from each image, total 444 for two observers from both PET and CT images. In CT texture analysis, 67 of 111 features in metastatic and non-metastatic patient groups showed statistically significant differences in both observers and 50 in PET texture analysis. As a result of the ICC analysis, CT 364 and PET 356 features were extracted. After Pearson correlation, 6, 7 and 8 features were selected with the wrapper-based recursive feature selection algorithm in CT, PET and combined group, respectively. The accuracy rates of SVM and LR models created with CT images were 81% for both, 75% and 68% for PET models, and 84% and 78% for the combined model, respectively. In the Wilcoxon test, the SVM model was found to be more successful than the LR model in combined and PET images (p = 0.027, 0.043). When LR models were compared with each other using the Friedman test, the model created with CT images was found to be more successful than PET images (p = 0.012). No significant difference was found between image sets in SVM models (p = 0.607). CONCLUSION: Lymph node metastasis is an important factor in the treatment and staging of malignant melanoma. In our study, machine-learning models for preoperative prediction of regional LN metastasis of malignant melanoma based on PET and CT based radiomic features were successfully developed and validated. According to the results obtained from our study, texture analysis and machine learning is promising in the diagnosis of metastatic lymph node in patients with malignant melanoma. Compared to SLNB, its non-invasiveness and reproducibility are important advantages. Compared to PET-CT examinations, CT provides high anatomical detail and no radionuclide exposure stands out as positive aspects. Our study is the only study in the literature to evaluate lymph node metastasis of malignant melanoma based on radiomics in PET and CT. Therefore, larger and deeper studies are needed to confirm the findings.