Skin cancer is the most common type of cancer worldwide. Its high incidence rate is a challenge for healthcare systems. Computer-aided diagnosis (CAD) systems have emerged as a promising tool to support skin cancer detection, particularly in remote regions. Recent advances in deep learning have enabled the development of multimodal models that integrate patient metadata with imaging data to improve diagnostic accuracy. However, these models may struggle to maintain high performance when dealing with missing or incomplete metadata. In this work, we propose an extension of the Metadata Processing Block (MetaBlock) architecture to address missing metadata in multimodal skin cancer classification. Our approach integrates a sentence embedding algorithm into the original method to generate dense vectors that capture the semantic relationships between the metadata features. We evaluate the proposed method on two different datasets: the PAD-UFES-20 and a new extended version of this dataset that is seven times larger. The proposed method outperforms the original MetaBlock architecture in all experiment scenarios, and obtains a balanced accuracy of up to 70. 2 2. 8 on the PAD-UFES-20 and 68. 2 1. 0 on the extended version with the ResNet-50 backbone. These results represents a step towards more robust multimodal models for skin lesion classification in real-world clinical scenarios. Finally, the new method is simple to implement and computationally efficient in terms of inference time and size, which makes it suitable to embed in real-world CAD systems.
Bouzon et al. (Mon,) studied this question.