Researchers at Queen Mary University of London have found a new way to improve the accuracy of thyroid cancer diagnoses using artificial intelligence (AI). The study used a special type of approach, called Generative Adversarial Networks (GANs), to create realistic images of thyroid tissue samples. These fake images were then used to train AI systems to better detect cancer pathologies in real-life samples, potentially leading to more accurate diagnoses.
Thyroid illustration.
Thyroid cancer is one of the top ten most common cancers globally. Misdiagnosis or overdiagnosis is a recognised issue with thyroid cancer, sometimes leading to unnecessary treatments or delayed care.
In the study, the GANs were trained using images from a small number of patients. Despite the limited data, the system learned to create highly realistic images of thyroid tissue, including different cancer subtypes. These images were then added to a training dataset for an AI model, which was used to predict the presence of cancer in new, unseen images. This method boosted the model’s accuracy and generalisability, particularly when dealing with difficult-to-classify tumour types.
When tested on data from three different sources, the enhanced model showed an improvement in precision by 7.45% and overall performance (measured by AUC) by 7.20%. It was especially good at identifying minority class images—subtypes of tumours that are hard even for trained pathologists to diagnose reliably.
Dr Eirini Marouli, Senior author and Associate Professor in Computational Biology at Queen Mary University of London said:
"One of the biggest challenges in AI implementation for medical imaging is getting enough data to train deep learning models. We wanted to see if we could use artificially generated data to fill this gap by creating synthetic images, and it was exciting to see how well it worked, even with a small dataset.”
Will Dee, first author and PhD student at Queen Mary University of London added: "Our findings suggest that AI models, supplemented with GAN-generated data during training, could help pathologists make more accurate diagnoses, especially for borderline cases. With more data, we could train even more powerful models that could improve clinical decision-making and patient outcomes".
Further information