Localized And Global Feature Integration for Efficient Bladder Cancer Detection Using Deep Learning

Abstract

Bladder Cancer (BC) is the prevalent urinary system cancer with abnormal cell growth in the bladder lining that, if not detected at an early stage, may cause life-threatening complications. Bladder cancer detection includes identifying and categorizing cancer and non-cancer cases through histopathological imaging methods. Making an accurate diagnosis is essential to initiate therapy on time and enhance patient outcomes. Challenges like the heterogeneity of tumor appearance, overlapping imaging features with benign conditions and imaging artifacts make the diagnosis challenging. Conventional approaches are based on invasive techniques like cystoscopy and manual image analysis, which are time-consuming and unpredictable. The present study focuses on developing a hybrid Deep Learning (DL) model that integrates local and global feature extraction methods for effective bladder cancer detection. The dataset utilized consists of 3045 histopathological images, 1435 of which are classified as healthy and 1610 as Urothelial Cell Carcinoma (UCC). Preprocessing techniques, including normalization and augmentation, are applied to enhance data quality and variability. The hybrid model utilizes Convolutional Neural Networks (CNN) for local feature representation and EfficientNetB0 for global feature representation, combining their outputs using dense layers in order to classify as "healthy" or "UCC". The suggested study makes a high-impact performance with 98.68% accuracy, 98.75% precision, 98.75% recall and an F1-score of 98.75%, outperforming all current approaches to bladder cancer detection. These outcomes denote the success of the hybrid model in overcoming bladder cancer diagnosis challenges, providing a robust and reliable solution to medical practice.