Evolutionary Neural Architecture Search for Bladder Cancer: A Multi-Objective Investigation

Abstract

Artificial intelligence holds significant promise for enhancing diagnostic accuracy and mitigating subjectivity in medical imaging, particularly for high-stakes tasks such as cancer detection. However, model performance frequently degrades when transitioned to real-world clinical scenarios. Employing a proprietary clinical dataset consisting of white-light cystoscopy videos acquired from the Amiens University Hospital, this study investigates the design of robust Convolutional Neural Networks (CNNs) through the Non-dominated Sorting Genetic Algorithm (NSGA-II). The search process was configured to automatically generate architectures that maximize both mean sensitivity and mean specificity within a 5-fold cross-validation framework. Search efficiency was evaluated using the HI. The evolved architectures were subsequently benchmarked against VGG16, ResNet50, and DenseNet121 models pretrained on ImageNet. Experimental results show a gradual increase in the HI from 0.60 to 0.67 during the initial phase, followed by a decline upon the inclusion of the third fold, suggesting a "generalization shock" before a final marginal recovery. These fluctuations coincided with a substantial reduction in genotypic diversity from 0.87 to 0.28, leading to premature convergence in the NSGA-II optimization. Consequently, the resulting CNN architecture remained below the ResNet50 baseline, highlighting that further refinement of optimization strategies and training hyperparameters is essential to bridge the gap between automated design and expert-crafted models.