Ai-Generated Clinical Image Using StyleGAN2-ADA for Improving Rectal Cancer Screening

Abstract

This research focuses on the improvement of classification of abnormal and normal rectum CT scan images in the context of colorectal cancer. Limited data availability poses a challenge as classification models, particularly deep learning models, require substantial amounts of data to achieve optimal performance. Furthermore, lack of understanding regarding CT image processing such as windowing in classification tasks leads to unoptimized diagnostic potential of medical imaging data. To address this issue, we employ a Generative Adversarial Networks (GANs) model to generate additional data using the existing dataset. Subsequently, we train the classification model to classify the rectum CT scan images as normal or abnormal by using the real data and utilizing the generated data to assess any improvements of the model. We also train the classification by using rectum CT scan images in different window settings such as bone window and tissue window to compare their performance. Our findings indicate that StyleGAN2-ADA effectively generates synthetic rectum CT scan images of normal and abnormal patients. Leveraging the generated images, we observe an enhancement in the classification performance up to 55 percents gains specifically on the smaller model, MobilenetV3. These results suggest the potential of data generation using GAN-based models to improve the accuracy and efficacy of classification models in the field of medical imaging. Overall, InceptionV3 is the best classification model to classify normal and abnormal CT scan images. Lastly, there is no significant classification performance difference observed between tissue and bone window image modalities, with both settings demonstrating similar classification accuracy. This research represents an initial step towards integrating AI into the Malaysian healthcare system to support healthcare professionals. With continued implementation, the Malaysian healthcare industry stands to benefit significantly from this technology, paving the way for a brighter future.