Synthetic training data generation from a single image for enhanced breast cancer diagnosis

Abstract

[en] According to the World Health Organisation (WHO), breast cancer is one of the cancer types with a high prevalence worldwide. Deep-learning based computeraided detection systems have shown promising potential in improving the curability and reducing mortality rates through early detection in mammography screening. Artificial Intelligence (AI) has become a popular tool in medicine, aiming to reduce costs and assist radiologists in decision-making processes. However, AI in cancer imaging presents significant challenges, including data access and privacy issues, as well as a scarcity of expert-annotated medical imaging. Motivated by these factors, this project aims to enhance the robustness and generalisability of breast cancer classification tools. The study focuses on obtaining a pre-biopsy result of suspicious areas in mammograms, providing a comprehensive assessment of lesion nature. It was observed that the classifier’s performance for the malignant class was inferior to that of the other classes, and the tightness of the annotation mask around the lesion significantly influenced the classifier’s performance. To improve the performance for malignant lesions, the study investigates data augmentation based in single image Generative Adversarial Network (SinGAN) to balance this underrepresented class. To the best of our knowledge, this project represents a novel investigation into the application of single-image generative models for breast cancer, addressing the challenge of expert annotation scarcity. Promising results were observed through the use of SinGAN-based data augmentation. The classification model, trained with SinGAN-augmented training data, demonstrated a higher area under the receiver operating characteristic (AUROC) for the malignant class (0.718 ± 0.044), compared to the same model without augmented data (0.677 ± 0.076). Furthermore, it was also identified an unexpected trend during the experiments. It was observed that using more SinGANs for data augmentation did not always result in a higher enhancement of performance. This project opens up new research possibilities through collaboration with healthcare experts. Its ultimate goal is to analyse and validate a mitigation strategy for improving robustness and, as such, trustworthiness of AI-based applications for adoption in the clinical workflow.