Глубокое обучение в выявлении раннего папиллярного рака мочевого пузыря на ограниченном наборе цистоскопических изображений

Aim. The aim of this study was to develop and evaluate the effectiveness of a convolutional neural network (CNN) in detecting papillary bladder cancer (PBC) using a limited set of cystoscopic images. Materials and methods. Twenty patients who underwent white light cystoscopy and histologically confirmed papillary bladder cancer were included in the study. The dataset included 125 images retrieved and marked by a urologist: 88 images were papillary tumors and 37 were healthy bladder wall tissue. 100 images were selected for training and 25 images were selected for validation. The U-net architecture and the CNN VGG16 model were used. A binary mask was manually created for each image based on the comments given by the urologist. Each image was additionally processed for model compatibility, with 224×224 pixel images as input to reduce the number of parameters. The dataset was augmented by applying vertical and horizontal turns, as well as random rotations. The following metrics were calculated: Dice coefficient, sensitivity, specificity, proportion of false positives and false negatives, accuracy, and area under the ROC curve. Results. The original data set yielded the following parameters: specificity 84.56%, sensitivity 82.18%, false positive rate 15.44%, false negative rate 17.82%, accuracy 76.40%, and a Dice coefficient 83.16%. For the augmented dataset, the following values were obtained: specificity: 82.99%, sensitivity: 82.70%, false positive rate 17.01%, false negative rate 17.30%, accuracy 74.72%, Dice coefficient – 82.82%. The area under the ROC curves was 92.93% for the original dataset and 91.69% for the augmented dataset. Conclusion. The CNN created in this study can detect signs of early PBC when analyzing cystoscopic images. The results of the study can be a starting point for developing new methods to diagnose PBC using deep learning technologies.