Advances in Robustness and Generalization in Computer Vision and Medical Imaging

The field of computer vision and medical imaging is moving towards developing more robust and generalizable models. Recent research has focused on improving model performance in the presence of adversarial attacks, distribution shifts, and limited training data. Techniques such as multi-task learning, hierarchical Bayesian models, and fusion of heterogeneous models have shown promise in addressing these challenges. Additionally, there is a growing interest in developing efficient and interpretable models for resource-constrained applications. Notable papers in this area include: VISAT, which introduces a benchmarking suite for evaluating model robustness in traffic sign recognition; CoMViT, which presents a compact and generalizable Vision Transformer architecture for medical image analysis; and PLUTO-4, which introduces a new generation of pathology foundation models that achieve state-of-the-art performance on various histopathology tasks. These advancements have the potential to transform real-world applications in autonomous driving, healthcare, and other fields.

Sources

VISAT: Benchmarking Adversarial and Distribution Shift Robustness in Traffic Sign Recognition with Visual Attributes

Hierarchical Bayesian Model for Gene Deconvolution and Functional Analysis in Human Endometrium Across the Menstrual Cycle

Fusion of Heterogeneous Pathology Foundation Models for Whole Slide Image Analysis

CoMViT: An Efficient Vision Backbone for Supervised Classification in Medical Imaging

Beyond ImageNet: Understanding Cross-Dataset Robustness of Lightweight Vision Models

STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology

LL-ViT: Edge Deployable Vision Transformers with Look Up Table Neurons

Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow

Scalable Evaluation and Neural Models for Compositional Generalization

PLUTO-4: Frontier Pathology Foundation Models

Improving Diagnostic Performance on Small and Imbalanced Datasets Using Class-Based Input Image Composition

Built with on top of