Alignment of Deep Neural Representations and Radiomics for Improved Kidney Tumor Classification

I-Chun Tsai, Hsi-Yu Ho, Huan-Yu Chen, Chun Chieh Weng, Jing-Tong Tzeng, Ching-Heng Lin, Chi-Chun Lee, Po-Chih Kuo

Keywords: Radiomics, Data Fusion, Deep Learning, Convolutional Neural Networks

Abstract: Preoperative diagnosis of kidney tumors, including the crucial differentiation between benign and malignant lesions and the grading of malignant tumors (low-grade vs. high-grade), is vital for treatment planning but challenging with traditional imaging alone. This study proposes a novel Radiomics-Guided Convolutional Neural Network (RGCNN) model that aims to enhance kidney tumor classification accuracy by integrating Computed Tomography (CT) images with quantitative radiomic features. Data preprocessing involved isotropic resampling and Connected Component (CC3D) cropping on datasets from three different sources (KiTS19, TCGA-KIRC, VGHTC) to standardize the tumor region. The RGCNN model not only utilizes Traditional Radiomics (TR) features but also extracts CNN-Based Radiomics (CR) features from its convolutional layer feature maps. Experimental results show that combining both CR and TR outperforms standalone CNN or TR-only models. The combined model achieved an AUC of 0.82 and an F1-Score of 0.89 in the benign-versus-malignant classification task, and an AUC of 0.74 and an F1-Score of 0.64 in the more challenging low-grade versus high-grade classification task. This study validates the significant potential of RGCNN as a non-invasive, preoperative diagnostic aid.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 48