Spectrum Lab

Vision loss remains a critical global health issue, with conditions such as diabetic macular edema and age-related macular degeneration often detected only after substantial retinal damage. Early retinal fluid accumulation is a key biomarker, and Optical Coherence Tomography (OCT) provides noninvasive, high-resolution retinal imaging essential for diagnosis. Accurate segmentation of retinal structures is crucial, yet recent conventional deep learning approaches often do not provide satisfactory performance because fluid regions occupy a small portion of the image, leading to inefficient computation and reduced accuracy. We propose CLAVIS, a novel plug-and-play pipeline that leverages CLAssification outputs to generate VIsualizations and guide Segmentation, generating more precise maps of the fluid regions. We leverage visualization techniques to target image regions that influenced the classification outcome. Experimental results demonstrate that this guidance strategy improves Dice score by $\mathbf{+ 0. 1 4}$ on the average and up to $\mathbf{+ 0. 1 8}$ for some pathologies.

Sparse signal recovery is an important problem in compressed sensing, signal/image processing, computational imaging and machine learning. The design of the sensing matrix is crucial in compressed sensing. While random Gaussian sensing matrices are relatively easier to construct, it is known that matrices constituting a tight frame provide superior reconstruction performance. We formulate the problem of sparse signal recovery using a new data-fidelity term that effectively "tightens" the sensing matrix, and incorporate the minimax concave penalty (MCP) instead of the ℓ1-norm for promoting sparsity. We carry out the optimization using a proximal gradient method and its Nesterov accelerated counterpart. Simulation results demonstrate that the proposed algorithms result in up to three times faster convergence and superior reconstruction accuracy over benchmark methods.

3D Gaussian Splatting (3DGS) has revolutionized novel view synthesis by offering excellent visual quality, real-time rendering and faster training. However, the infinite support and inherent smoothness of Gaussian kernels limit their ability to reconstruct fine geometric details and high-frequency texture, while also introducing domain-clipping artifacts near scene boundaries. We introduce SplineSplat, a novel approach that replaces Gaussian kernels with compactly supported and scalable polynomial B-spline bases, specifically, linear, quadratic, and cubic splines. B-Splines offer improved spatial localization, greater representational flexibility, and efficient evaluation through piecewise-polynomial definitions. We utilize fast spline computation strategies that enable significant training acceleration without sacrificing rendering quality or speed. Evaluations on the Mip-NeRF 360 static dataset and NVIDIA dynamic dataset demonstrate that SplineSplat matches the reconstruction fidelity of 3DGS while reducing training time by up to 18\%, making it an effective kernel for high-performance radiance-field representations.