Comparative Analysis of CNN Architectures for High-Accuracy Subtype Classification of Renal Cell Carcinoma Using CECT

Goal: Objective: Renal Cell Carcinoma (RCC) is a heterogeneous cancer with multiple histological subtypes that require individualized clinical management. Once implemented, timely and appropriate subtype delineation from imaging data can lead to improved diagnostic accuracy and intervention decisions. Methods: Study compares 3 convolutional neural networks (CNN) architectures, AlexNet, GoogleNet, VGG16, Resnet18 and Densenet to determine which approach has the best performance for RCC subtype classification of Contrast-Enhanced computed tomography (CECT) images. The original dataset of 590 images from 90 patients was augmented using anatomically constrained augmentation (rotation ±5° and 2D scaling) to generate a total of 3469 unique images, with allocation of training, validation, and testing sets was 70:15:15, respectively. Each architecture was fine tuned to employ an optimized learning rate, dropout regularization, and L2 penalties to minimize overfitting. Results: The quantitative metrics for each architecture, displayed in Table IV, show that VGG16 achieved the maximum level of 99.21% accuracy but demonstrated some levels of overfitting behavior at the lower learning rates. GoogleNet achieved commendable accuracy levels of 97.67% accuracy after epoch optimization, but performance levels were highly sensitive across the learning rate. AlexNet, in comparison demonstrated the most stable convergence rates and achieved strong levels of generalization within all three large learning rates, achieving levels of 98.84% accuracy, and processing times were far lower than for either of the other networks. Conclusion: this work demonstrates that GoogleNet is a reliable and cost efficient implementation of a CNN for a clinical setting that enables rapid accurate clinical decision making from imaging data for the specific subtype of RCC. Significance: These findings reinforce the potential of tailored deep learning frameworks for early RCC subtype classification and encourage further validation on larger, multi-institutional datasets to enhance clinical applicability.