A Generative Adversarial Network (GAN)-based Three-dimensional Rectal Cancer Pathological Complete Response Prediction Framework

Abstract

Abstract Colorectal cancer affects the health of the global public, and the increasing proportion of cases has attracted widespread attention. This phenomenon has made the treatment of colorectal cancer an inevitable topic in the global medical community, and has sparked interest in using deep learning models for early detection and diagnosis of colorectal cancer. This study proposes a method based on Three-dimensional (3D) Magnetic Resonance Imaging (MRI) data to predict the complete pathological remission of rectal cancer patients. To improve prediction accuracy, we employ an improved Deep Convolutional Generative Adversarial Network (DCGAN) for data augmentation and optimize the 3D network with different attention modules. Specifically, we employed a DCGAN generator for data augmentation. Instead of using deconvolution operations as in the DCGAN generator, we utilized upsampling and convolution operations to diminish the impact of "artifacts" on the generated images. Additionally, we enhanced the image quality by utilizing an improved AlexNet-based discriminator architecture. Furthermore, we utilize the Convolutional Block Attention Module (CBAM) for feature extraction and capturing spatial and channel information. The experimental results of this study demonstrate significant improvements in accuracy, specificity, and sensitivity through the application of data augmentation and attention mechanisms. In detail, the accuracy is improved to 0.778, specificity to 0.796, and sensitivity to 0.754. Compared to the baseline network, these values have increased by 8.8%, 9.9%, and 9.1% respectively. These findings indicate that the method we propose offers a potential tool for doctors to avoid unnecessary surgical procedures.