Differentiate Xp11.2 Translocation Renal Cell Carcinoma from Computed Tomography Images and Clinical Data with ResNet-18 CNN and XGBoost

Abstract

This study aims to apply ResNet-18 convolutional neural network (CNN) and XGBoost to preoperative computed tomography (CT) images and clinical data for distinguishing Xp11.2 translocation renal cell carcinoma (Xp11.2 tRCC) from common subtypes of renal cell carcinoma (RCC) in order to provide patients with individualized treatment plans. Data from 45 patients with Xp11.2 tRCC from January 2007 to December 2021 are collected. Clear cell RCC (ccRCC), papillary RCC (pRCC), or chromophobe RCC (chRCC) can be detected from each patient. CT images are acquired in the following three phases: unenhanced, corticomedullary, and nephrographic. A unified framework is proposed for the classification of renal masses. In this framework, ResNet-18 CNN is employed to classify renal cancers with CT images, while XGBoost is adopted with clinical data. Experiments demonstrate that, if applying ResNet-18 CNN or XGBoost singly, the latter outperforms the former, while the framework integrating both technologies performs similarly or better than urologists. Especially, the possibility of misclassifying Xp11.2 tRCC, pRCC, and chRCC as ccRCC by the proposed framework is much lower than urologists. KEYWORDSResNet-18 CNN; XGBoost; computed tomography; TFE3; renal cell carcinoma Materials and Methods Patient EnrollmentThe complete CT imaging data and postoperative pathological data of patients with Xp11.2 tRCC diagnosed in the Nanjing Drum Tower Hospital were retrospectively analyzed from January 2007 to December 2021. 45 cases of Xp11.2 tRCC with complete clinicopathological data were enrolled in the study, including complete medical histories, clinical information, and pathological data. According to both tumor size and pathological stage, the matched patients included 45 with ccRCC, pRCC, or chRCC, respectively. The study was approved by the Institutional Review Board of Nanjing Drum Tower Hospital. Inclusion criteria:1. Patients underwent radical or partial nephrectomy. 2. Preoperative CT examination of renal tumors and complete examination records. 3. The reviewed pathological diagnosis was consistent with the initial postoperative diagnosis and the pathological diagnosis of Xp11.2 tRCC was validated by the next generation sequence. 4. These patients were without cardiovascular and cerebrovascular diseases and chronic kidney disease that affected the blood supply of the kidney. 5. These patients had a unilateral single renal tumor.

SNSVM: SqueezeNet-Guided SVM for Breast Cancer Diagnosis

“…SqueezeNet [14,36] is chosen as the base architecture for this project due to its small size and low number of parameters. Compared to larger models like AlexNet [37][38][39], VGG [40,41], or ResNet [42][43][44], SqueezeNet has a significantly smaller number of parameters, making it easier to train on limited computational resources. Table 1 compares the number of parameters of various deep learning models.…”