Analysis and Generation in Domains of Limited Data: From Illusions to X-Ray Imaging

While most machine learning techniques rely on large datasets, such data is not readily available in all domains. Our work began by analyzing and generating visual illusions, where a dedicated dataset is not helpful, and later shifted focus to image-to-text generation. This work explores innovative solutions for image-to-text generation in data-limited scenarios across natural and medical imaging domains. For natural images, we introduce CLID, a novel approach for length-controlled image captioning, particularly suited for generating long, descriptive captions. In the medical imaging domain, specifically chest X-rays, privacy restrictions and high labeling costs limit dataset availability. To address this, we propose two approaches, MedCycle and MedRAT, for generating medical reports in an unpaired setting, where images and reports are unmatched and originate from different sources. These advancements offer promising solutions for image analysis in data-scarce environments.

Ph.D. Under the supervision of Prof. Ayellet Tal.