The recent success and exponential use of artificial intelligence (AI) in medical imaging is largely due to the successful application of deep learning to labeled big data. Notably, convolutional neural networks (CNNs) have great capabilities in image recognition tasks [1]. Since 2018, a total of eleven data challenges have been organized by the French Society of Radiology (SFR), with two on ultrasound images, three on magnetic resonance imaging (MRI) and six on computed tomography images [2], [3], [4]. The previous data challenges led by the SFR demonstrated high performances of CNNs in lesion detection, segmentation and classification tasks, applied to cervical lymphadenopathies [5], pulmonary nodules [6] or breast nodules [7].

However, the performance of CNNs may be limited when data is sparse, as it is the case for a rare disease. In such situations, a model can be trained to the point of perfectly predicting labels on the training data but poorly on independent test data, reflecting an overfitted and poorly generalizable model [8]. In that context, generative adversarial networks (GANs) have recently gained great interest. GANs have the potential to increase the number of training images by creating fake images that look like real images [9].

The 2021 edition of the Artificial Intelligence Data Challenge organized by the SFR together with the Centre National d'Etudes Spatiales (CNES) focused on the macrotrabecular-massive (MTM) subtype of hepatocellular carcinoma (HCC), a rare and aggressive type of primary liver cancer with poor prognosis [10,11]. MTM-HCC displays suggestive imaging features on contrast-enhanced MRI including substantial necrosis and diffuse hypovascular component [12,13].

The purpose of this data challenge was to create a synthetic dataset of 1000 MTM-HCC cases from a limited number of real cases using GAN-based data augmentation techniques.