Seminar AI4Science: Unsupervised Domain Adaptation for Optoacoustic Imaging: Leveraging Unpaired Data for Improved Results

With limited data or annotations, a common practice in AI is to build upon or exploit models pretrained on large-scale natural image datasets and/or generic natural language datasets. However, in domain sciences, such as optoacoustic imaging, it can be challenging to benefit from such pretrained models. Optoacoustic imaging involves the reconstruction of tissue images from acoustic signals generated by a light source. As a novel non-invasive tomographic imaging modality, optoacoustic imaging has a strong potential. In this talk, I describe:
(i) optoacoustic imaging in simple terms,
(ii) some of the challenges that can be tackled with data science techniques,
(iii) supervised translation with limited-quality paired datasets,
(iv) improvements we achieved with unsupervised image-to-image translation using unpaired datasets (e.g., CycleGAN).
This is a joint work with the Multiscale Functional and Molecular Imaging Lab at ETH Zurich.