On Urinary Bladder Cancer Diagnosis: Utilization of Deep Convolutional Generative Adversarial Networks for Data Augmentation

Abstract

Urinary bladder cancer is one of the most common urinary tract cancers. Standard diagnosis procedure can be invasive and time-consuming. For these reasons, procedure called optical biopsy is introduced. This procedure allows in-vivo evaluation of bladder mucosa without the need for biopsy. Although less invasive and faster, accuracy is often lower. For this reason, machine learning (ML) algorithms are used to increase its accuracy. The issue with ML algorithms is their sensitivity to the amount of input data. In medicine, collection can be time-consuming due to a potentially low number of patients. For these reasons, data augmentation is performed, usually through a series of geometric variations of original images. While such images improve classification performance, the number of new data points and the insight they provide is limited. These issues are a motivation for the application of novel augmentation methods. Authors demonstrate the use of Deep Convolutional Generative Adversarial Networks (DCGAN) for the generation of images. Augmented datasets used for training of commonly used Convolutional Neural Network-based (CNN) architectures (AlexNet and VGG-16) show a significcan performance increase for AlexNet, where AUCmicro reaches values up to 0.99. Average and median results of networks used in grid-search increases. These results point towards the conclusion that GAN-based augmentation has decreased the networks sensitivity to hyperparemeter change.

Keywords: AlexNet; VGG16; data augmentation; deep convolutional generative adversarial networks; urinary bladder cancer.

Figure 1

Dataflow diagram of Convolutional Neural Network-based (CNN) utilization in urinary bladder cancer diagnosis.

Figure 2

Representation of all four cystoscopy dataset classes.

Figure 3

The process of Generative Adversarial Networks (GAN) use for a single image type is shown. Data is split into training and testing sets, and training set is used to generate additional images. Generated images are mixed with existing dataset and used for training, generating models which are finally evaluated on the testing set.

Figure 4

Illustration of generation process of generation and discrimination of images.

Figure 5

Comparison of images representing healthy mucosa generated by GAN executed for: 100 (a), 250 (b) 500 (c) and 1000 (d) epochs respectively.

Figure 6

Illustration of a cross-validation procedure adapted for GAN-based augmentation.

Figure 7

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of AlexNet trained with images generated by GAN in 100 epochs.

Figure 8

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of AlexNet trained with images generated by GAN in 250 epochs.

Figure 9

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of AlexNet trained with images generated by GAN in 500 epochs.

Figure 10

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of AlexNet trained with images generated by GAN in 1000 epochs.

Figure 11

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of VGG-16 trained with images generated by GAN in 100 epochs.

Figure 12

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of VGG-16 trained with images generated by GAN in 250 epochs.

Figure 13

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of VGG-16 trained with images generated by GAN in 500 epochs.

Figure 14

Median, average $\overline{AU{C}_{micro}}$ and standard deviation of $\overline{AU{C}_{micro}}$ and $\sigma \left(AU{C}_{micro}\right)$ for the case of VGG-16 trained with images generated by GAN in 1000 epochs.

Figure 15

Comparison of achieved results.