Data Augmentation: A Combined Inductive-Deductive Approach Featuring Answer Set Programming.

This article presents a novel framework for image data augmentation in the biomedical domain that integrates Answer Set Programming (ASP), a declarative logic programming formalism, with deep learning (DL) techniques. The approach begins with a limited set of labeled images and uses ASP to generate new semantically labeled images by encoding domain-specific knowledge and constraints, ensuring realistic and customizable image content. These labeled images are then transformed into photo-realistic images using DL-based methods, specifically Semantic Image Synthesis With Spatially-Adaptive Normalization (SPADE). The framework was tested on a Laryngeal Endoscopic Images dataset, demonstrating promising results in generating high-quality synthetic images that respect medical domain knowledge, outperforming traditional Generative Adversarial Networks (GANs) and Vision Transformers (ViTs) in fidelity and diversity metrics. This hybrid inductive-deductive approach highlights the potential of combining declarative knowledge representation with DL for controlled and explainable biomedical image augmentation.