A Deep Learning-Based Radiomics Model for Differentiating Benign and Malignant Renal Tumors

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Renal tumors require accurate classification into benign and malignant types for appropriate treatment.
• Computed Tomography (CT) is a key imaging modality for renal tumor assessment.
• Transfer learning offers a promising approach to enhance diagnostic accuracy in medical image analysis.

Purpose of the Study:

• To evaluate the efficacy of transfer learning using CT images for classifying renal tumors.
• To develop and assess patient-level models to improve classification accuracy compared to image-level models.

Main Methods:

• A dataset of 192 renal tumor cases underwent enhanced CT examination and pathological diagnosis.
• The InceptionV3 deep learning model, pre-trained on ImageNet, was adapted for renal tumor classification.
• Image-level models (Slice, ROI, RBR) and subsequent patient-level models were constructed and evaluated using ROC analysis and five-fold cross-validation.

Main Results:

• Image-level models showed varying performance, with the ROI dataset achieving higher accuracy (ACC=0.97, MCC=0.93) than Slice (ACC=0.69, MCC=0.45) and RBR (ACC=0.93, MCC=0.85).
• Patient-level models outperformed image-level models, demonstrating a 2%-5% increase in accuracy (ACC) on RBR and Slice datasets.
• The study highlights the superior performance of models utilizing region of interest (ROI) and patient-level data aggregation.

Conclusions:

• Deep learning, particularly transfer learning, is a viable tool for classifying benign and malignant renal tumors from CT imaging.
• Patient-level models show potential for enhanced diagnostic accuracy, likely by leveraging volumetric (3D) data insights.
• Further development of patient-level models may improve the clinical utility of AI in renal tumor diagnosis.