Bladder Cancer Recurrence Prediction: Using CT Scans and Convolutional Neural Networks with Transfer Learning (Preprint)

<sec> <title>BACKGROUND</title> Deep convolutional neural networks have been applied to predict bladder cancer recurrence by analysing CT scans. However, such models suffer from a limited number of images collected from hospitals, leading to unstable performance. The size of medical datasets is limited by the complexity and high cost of large-scale studies, making it difficult to obtain rich training data in medical research. </sec> <sec> <title>OBJECTIVE</title> In this study, we aim to investigate 13 mainstream deep convolutional neural networks and build a pipeline of bladder cancer recurrence risk classification using CT scans and pre-trained deep learning convolutional neural networks. </sec> <sec> <title>METHODS</title> A five-year follow-up dataset from 101 patients with bladder cancer was collected from the Department of Urology at Xuzhou Central Hospital, Jiangsu Province, China. The CT scans of each patient were extracted from the dataset and formed an image set of 492 scans. 5-fold cross-validation were used to assess the performance and general error estimation of deep learning models. 13 mainstream deep convolutional neural networks were built to predict bladder cancer's recurrence risk level. Model performance was assessed by accuracy, precision, recall, F1-score and Area Under the Curve (AUC). </sec> <sec> <title>RESULTS</title> The pre-trained VGG16 had the best accuracy of 0.954 compared to the other 12 pre-trained models. The model also had a precision of 0.945, recall of 0.946, and F1-score of 0.943. The AUC for short-term recurrence was 0.978, for medium-term recurrence was 0.981, and for long-term recurrence was 0.975. </sec> <sec> <title>CONCLUSIONS</title> Transfer learning provided a new way to deal with bladder cancer recurrence risk classification with deep learning and small-scale dataset. A pre-trained VGG16 performed the best in predicting the risk of bladder cancer recurrence using CT scans. </sec>