Advances in Robustness and Uncertainty in Medical Imaging

The field of medical imaging is moving towards developing more robust and reliable models, with a focus on addressing the challenges of distributional shifts, uncertainty, and heterogeneity in data. Researchers are exploring new approaches to improve the performance of deep learning models in real-world deployment, including the use of probabilistic guarantees, distributionally robust training, and graph-radiomic learning. These innovations have the potential to enhance the accuracy and trustworthiness of medical imaging models, particularly in safety-critical applications. Notable papers in this area include: Probabilistic Conformal Coverage Guarantees in Small-Data Settings, which introduces a plug-and-play adjustment to conformal significance levels to provide probabilistic guarantees. NeuroRAD-FM: A Foundation Model for Neuro-Oncology with Distributionally Robust Training, which develops a neuro-oncology specific foundation model with a distributionally robust loss function to improve cross-institution generalization. Graph-Radiomic Learning Descriptor to Characterize Imaging Heterogeneity in Confounding Tumor Pathologies, which presents a new descriptor for characterizing intralesional heterogeneity on clinical MRI scans. Probabilistic Runtime Verification, Evaluation and Risk Assessment of Visual Deep Learning Systems, which proposes a novel methodology for verifying, evaluating, and assessing the risk of deep learning systems. Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network, which presents a novel representation learning framework that aligns embeddings of cells subjected to the same perturbation across different wells. Anomaly Detection by Clustering DINO Embeddings using a Dirichlet Process Mixture, which leverages informative embeddings from foundational models for unsupervised anomaly detection in medical imaging.

Sources

Probabilistic Conformal Coverage Guarantees in Small-Data Settings

NeuroRAD-FM: A Foundation Model for Neuro-Oncology with Distributionally Robust Training

Comparing Computational Pathology Foundation Models using Representational Similarity Analysis

Influence of Classification Task and Distribution Shift Type on OOD Detection in Fetal Ultrasound

Graph-Radiomic Learning (GrRAiL) Descriptor to Characterize Imaging Heterogeneity in Confounding Tumor Pathologies

Probabilistic Runtime Verification, Evaluation and Risk Assessment of Visual Deep Learning Systems

Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network

Anomaly Detection by Clustering DINO Embeddings using a Dirichlet Process Mixture

Built with on top of