Deep learning has emerged as a powerful tool for analyzing spatiotemporal medical imaging data, offering the potential to significantly reduce costs associated with the workflows in imaging modalities such as dynamic positron emission tomography (dPET). Using the example of deep learning to predict the arterial input function (AIF), a critical component for kinetic modeling that enables in vivo quantification of tracer uptake and glucose metabolism, I will address some key challenges in this domain such as the scarcity of labeled training data, a common limitation in biomedical applications. I will cover strategies to address this issue, e.g by integrating physics-based constraints, leveraging domain knowledge to enhance model performance and generalizability.

In the pursuit of photorealism, computer graphics has traditionally relied on dense data structures, high-resolution textures, exhaustive BRDF tables, and massive photon maps, and suffered from the "curse of dimensionality”. Simulating the infinite bounces of light scattering in a complex scene or capturing the intricate appearance and spatially-varying reflectance of real-world materials (SVBRDFs) is prohibitively slow due to the computational complexity and the number of measurements, samples required. This presentation gives an overview of the transition from brute-force sampling to sparsity-aware capture and synthesis. By leveraging the fact that both light transport and material properties are inherently sparse in the appropriate basis, we can achieve cinematic quality with a fraction of the data.

I will present ongoing research on image analysis that moves beyond the traditional visual spectrum, motivated by the simple observation that cameras fail precisely in many of the environments where autonomy is most needed. Fog, dust, smoke and darkness routinely degrade optical sensing and challenge state-of-the-art perception systems. Recent advances in radar sensor hardware are changing this landscape. Modern radars now provide so-called 4D sensing, capturing not only range, azimuth, and elevation, but also radial velocity at high spatial and temporal resolution. This additional structure enables image analysis for computer vision and autonomous mobility even when vision breaks down.
I will outline our work on radar-centric perception, highlighting how 4D radar data can be interpreted for robust scene understanding and navigation in degraded visibility. The broader message of the talk is that reliable autonomy in the real world requires perception systems that are designed for failure modes, not just for ideal conditions, and that expanding the notion of "imagery" beyond light is a key step in that direction.