Bilgisayarli tomografi görüntülerinden makine ögrenmesi ileberrak hücreli böbrek karsinomun von hippel lindau (VHL)ve polybromo-1 (PBRM1) mutasyonlarinin ve evrelerinintahmin edilmesi

RCC is the most prevalent renal malignancy and ccRCC is the most common subtype of RCC. It is reported that the prognosis has a strong association with VHL alteration, and VHL mutation plays a role as a predictive and survival marker for ccRCC. It is also reported that PBRM1 gene has great potential to identify ccRCC and a critical role in ccRCC progression. It is the second most common alteration in ccRCC. Moreover, available treatment opportunities are mostly related to stage information. More than 50% of patients with early-stage RCC are cured, but the treatment options are limited in stage 3 and 4. Therefore, early diagnosis is major for the patients. The commonly used diagnosis method, namely biopsy, always has the potential to devastate the patient emotionally, damage the healthy tissue, or spread the tumor. In addition, studies of ccRCC indicate that there is a correlation between cancer CT imaging features and gene expression (radiogenomics). We hypothesized that from quantitative 2D CT images via one slice with the biggest tumor, both VHL and PBRM1 mutations and stages can be predicted with accuracy using machine learning algorithms. TCGAKIRC data were collected and divided according to specific gene mutations and stages. The tumor was segmented by an expert radiologist. After feature extraction, feature selection was performed. Finally, classification was done by using CL and ANN on Matlab. Our results showed that Fine Gaussian SVM model is able to predict VHL and NON-VHL data with 68.6%, k-NN with Random Subspace model is able to predict PBRM1 and NON-PBRM1 with 84.9% ,and ANN predicted stages with 91.90% accuracies. From this study, it appears that ML-based quantitative 2D CT analysis using one slice for each patient is a feasible and potential method for predicting the status of VHL and PBRM1 mutations and stages of patients with ccRCC.