BCDNet: GoogLeNet based Bladder Cancer Detection model from Urinary Cytological Images

In the realm of medical image processing, computational cytology is a crucial, quickly evolving, yet difficult topic that analyses the digitised cytology images using computer-aided technologies for bladder cancer screening. Both the incidence and mortality of bladder cancer have increased worldwide. This study intends to develop a deep learning based bladder cancer detection model from urinary cytological images for early detection of bladder cancer. The model, called BCDNet: a bladder cancer detection network, inspired form GoogLeNet is trained on voided urine cytological samples of 115 individuals, of which 59 are confirmed cases of Urothelial Cell Carcinoma and the remaining 56 are benign. The input images are first denoised and background removal is applied. These preprocessed cytology images are used to refine the GoogLeNet model in the ratio of 60:20:20, respectively for training, validation and testing. The proposed model is able to achieve and surpass the low sensitivity and low specificity of many of the existing state-of-the-art techniques in detecting the low grade tumors, with a classification accuracy of 95%, sensitivity 81%, and specificity of 91%. It is evident from the experimental results that BCDNet will provide clinicians a non-invasive and economic solution for the early detection of bladder cancer.