Diagnostic Classification of Cystoscopic Images Using Deep Convolutional Neural Networks

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Cystoscopic findings can be challenging to interpret, especially for less experienced clinicians.
• Computer-aided diagnosis (CAD) tools, particularly those using deep learning, offer a potential solution for diagnostic classification.

Purpose of the Study:

• To evaluate the efficacy of deep convolutional neural network (CNN) models for the diagnostic classification of cystoscopic images.
• To compare the performance of various CNN architectures and novel filter size concepts in identifying urologic findings.

Main Methods:

• A dataset of 479 patient cases with 44 urologic findings was utilized.
• Images underwent preprocessing, including color normalization, histogram equalization, rotation, and flipping, resulting in 18,681 images.
• Several CNN models (ResNet50, VGG-19, VGG-16, InceptionV3, Xception) were developed and evaluated using F1 scores, alongside two proposed CNN concepts.

Main Results:

• The Xception-based CNN model achieved the highest F1 score of 99.52%.
• ResNet50-based and harmonic-series concept models also demonstrated high performance (99.48% and 99.45% F1 scores, respectively).
• All cancer lesions were correctly identified; minor misclassifications occurred with bladder stones and bladder diverticulum images.

Conclusions:

• Deep learning models demonstrate significant potential for accurate diagnostic classification of cystoscopic images.
• Artificial intelligence-aided cystoscopy holds promise for integration into clinical practice and potential expansion to other endoscopic applications.