Hipofraksiyone Stereotaktik Radyocerrahi ile Tedavi Edilmiş Uveal Melanom Hastalarında Tedavi Öncesi Manyetik Rezonans Görüntüleme’den Elde Edilen Radyomiks Parametrelerinin Lokal Nüks ve Metastaz Gelişmesini Öngörmedeki Başarısının Araştırılması

Abstract

The annual incidence of uveal melanoma is about 6 per million population per year, yet it is the most common primary intraocular malignant tumor in adulthood. Hypofractionated stereotactic radiotherapy has recently emerged as an efficient alternative radiotherapy technique in the treatment of uveal melanoma due to its comparable effectiveness to other radiotherapy techniques, lower cost, lack of requirement for surgical intervention and easy accessibility. However, approximately one-quarter of cases experience local recurrence and 20% develop metastases within five years. In addition to the AJCC TNM staging system, molecular and genetic analyses requiring tissue diagnosis are currently used to predict prognosis in uveal melanoma. Studies on radiomics, which involve extracting multidimensional data from medical images, have shown to provide useful prognostic biomarkers. This study aims to investigate the value of radiomics features extracted from pre-treatment MRI images in predicting local recurrence and distant metastasis in patients with uveal melanoma treated with stereotactic radiosurgery, and to obtain potential biomarkers that could contribute to personalized medicine approaches. Among the 88 patients included in the study, local recurrence was detected in 16 patients and metastasis was observed in 16 patients within the four-year follow-up period, Two researchers independently performed manual segmentation of uveal melanoma lesions, resulting in a total of 2562 radiomics features, with 854 from each of the T1WI, FS-T2WI, and contrast-enhanced FS-T1WI sequences. After excluding features with a correlation coefficient below 0.80 between the two observers, 90 features for T1WI, 164 for FS-T2WI, and 91 for contrast-enhanced FS-T1WI were selected. Cases were randomly divided into a training set (70%) and a test set (30%). Feature selection for the local recurrence model was performed via elastic net regression algorithm which selected seven features including 4 from FS-T2WI and 3 from T1WI. The model developed using the Naive Bayes algorithm showed a performance of 92% accuracy (95% CI: 0.75, 0.99), 100% sensitivity, 91.6% specificity, 60% positive predictive value, and 100% negative predictive value in the test set. For predicting metastasis, five features, all from FST2WI were selected. The model created using the Naive Bayes algorithm achieved a performance of 81% accuracy (95% CI: 0.62, 0.93), 33% sensitivity, 87.5% specificity, 25% positive predictive value, and 91% negative predictive value in the test set. These results suggest that a radiomics-based machine learning model derived from MRI can be used as a biomarker to predict local recurrence after SRS treatment and could potentially contribute to personalized medicine approaches. However, for this model to be clinically applicable, it needs to be validated on multicenter datasets. The relatively low performance of the model in predicting metastasis may be due to overfitting, a common problem in machine learning. To overcome this, further studies with larger cohorts with balanced distribution of classes (metastasis present/absent) may be beneficial.