Introduction: 

Artificial intelligence (AI) has emerged as a powerful tool for enhancing diagnostic accuracy and reducing subjectivity in medical imaging[1], [2]. In the case of bladder cancer, the diagnosis relies mostly on bladder endoscopy, a tool used for the visualisation of the bladder's internal lining, the detection and the characterisation of potential lesions. However, its performance is imperfect and is highly dependent on the practitioner's experience [3], [4]. AI offers a great potential for the development of a diagnostic compagnon tool to endoscopy that standardises bladder cancer diagnosis.

Aim: 

The main purpose of this study is the development of a neural network architecture from scratch fine-tuned on for the detection of bladder cancer using a proprietary data set of white light cystoscopy images.

Methodology:

Between February 2022 and September 2025, 171 patients (265 videos) underwent a transurethral resection (TUR) of initial or recurrent suspected bladder tumor. In this study, a total of 82 videos were randomly selected with fifty-eight videos presenting cancerous lesions and 24 videos for control. From each video, 400 images were extracted with low quality images manually removed and 20 frames per video randomly selected for the final dataset.

We used the non-dominated genetic algorithm (NSGA-II) [5] for the bi-objective maximisation of a DenseNet-type architecture's sensitivity and specificity. These metrics were evaluated on a randomly predefined 80:20 train-test split stratified per video after training the model for 20 epochs. Using this algorithm, we optimised a set of parameters determining the architectural features of the model, some crucial training hyperparameters and some data-augmentation features. The optimisation process was carried on over 20 generations and the resulting pareto font for each generation was evaluated using the hypervolume index. Finally, the performance of the resulting model was then compared in a stratified-per-video 5-fold cross validation setting to other state of the art pretrained convolutional neural network (CNN) models.

Results:

During the NSGA-II process, the hypervolume indicator of the pareto front at each generation increased significantly from 0.72 to 0.93 showcasing a substantial enhancement in the quality and diversity of solutions. From the final pareto front, one architecture (NSGA-II-CNN) was able to balance both the objective metrics with a sensitivity of 87.9% and specificity of 89.0%. 

When compared to several pretrained convolutional neural networks (CNN) (VGG16 [6], ResNet50, ResNet151 [7], DenseNet121, DenseNet201 [8]) in the 5-fold cross validation settings, the NSGA-II-CNN showed superior performance stability a stronger robustness to overfitting. By the twentieth epoch, the performance of NSGA-II-CNN ranged between 0.6 and 0.8 for both sensitivity and specificity while ResNet50, DenseNet121, and DenseNet201 exhibited high variability in sensitivity across the 5 folds. On the other hand, ResNet151 was extremely unstable when it came to specificity. 

Conclusion:

The results of this study showcases the capability of genetic algorithms at designing tailored neural network architectures for medical diagnosis tasks. Our results demonstrated the capacity of NSGA-II at finding CNN architectures that balance specificity and sensitivity that surpasses some fine-tuned pretrained neural networks. 

While these results are promising, further studies need to be conducted on a larger and more diverse dataset through the integration of more videos. Further validations in a multi-center setting is also necessary to confirm robustness and reproducibility. Incorporating some interpretability techniques can also be interesting to confirm the models ability at extracting discriminative features and offer support to clinical decision-making.

Bibliography:

[1] Chen ZH, Lin L, Wu CF, Li CF, Xu RH, Sun Y. Artificial intelligence for assisting cancer diagnosis and treatment in the era of precision medicine. Cancer Commun (Lond). 2021 Nov;41(11):1100-1115. doi: 10.1002/cac2.12215. Epub 2021 Oct 6. PMID: 34613667; PMCID: PMC8626610.

[2] Shimizu H, Nakayama KI. Artificial intelligence in oncology. Cancer Sci. 2020 May;111(5):1452-1460. doi: 10.1111/cas.14377. Epub 2020 Mar 21. PMID: 32133724; PMCID: PMC7226189.

[3] Guldhammer CS, V´asquez JL, Kristensen VM, Norus T, Nadler N, Jensen JB, Azawi N. Cystoscopy Accuracy in Detecting Bladder Tumors: A Prospective Video-Confirmed Study. Cancers (Basel). 2023 Dec 28;16(1):160. doi: 10.3390/cancers16010160. PMID: 38201586; PMCID: PMC10777997.

[4] Zhu CZ, Ting HN, Ng KH, Ong TA. A review on the accuracy of bladder cancer detection methods. J Cancer. 2019 Jul 8;10(17):4038-4044. doi: 10.7150/jca.28989. PMID: 31417648; PMCID: PMC6692607.

[5] K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, ”A fast and elitist multiobjective genetic algorithm: NSGA-II,” in IEEE Transactions on Evolutionary Computation, vol. 6, no. 2, pp. 182-197, April 2002, doi:10.1109/4235.996017

[6] SIMONYAN, Karen et ZISSERMAN, Andrew. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.

[7] HE, Kaiming, ZHANG, Xiangyu, REN, Shaoqing, et al. Deep residual learning for image recognition. In : Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-778.

[8] HUANG, Gao, LIU, Zhuang, VAN DER MAATEN, Laurens, et al. Densely connected convolutional networks. In : Proceedings of the IEEE conference on computer vision and pattern recognition. 2017. p. 4700-4708.