Paper Info
Title
Does your dermatology classifier know what it doesn’t know? Detecting the long-tail of unseen conditions
Abstract
•We propose a novel hierarchical outlier detection (HOD) loss, and show that this outperforms existing outlier exposure based techniques for detecting OOD inputs.•We introduce a near-OOD benchmarking framework and the key design choices needed for proper validation of OOD detection algorithms.•We demonstrated the added utility of the novel HOD loss in the context of multiple different state-of-the-art representation learning methods (self-supervised contrastive pre-training based SimCLR and MICLe). We also show the OOD detection performance gains on large scale standard benchmarks (ImageNet and BiT model pre-trained on a large-scale JFT dataset).•We propose to use a diverse ensemble with different representation learning and objectives for improved OOD detection performance. We demonstrate its superiority over vanilla ensembles and performed analysis investigating how diversity aids in better OOD detection performance.•We propose a cost-weighted evaluation metric for model trust analysis that incorporates the downstream clinical implications to aid assessment of real-world impact.
Year
DOI
Venue
2022
10.1016/j.media.2021.102274
Medical Image Analysis
Keywords
DocType
Volume
Deep learning,Dermatology,Ensembles,Long-tailed recognition,Out-of-distribution detection,Outlier exposure,Representation learning
Journal
75
ISSN
Citations 
PageRank 
1361-8415
0
0.34
References 
Authors
0
21
Name
Order
Citations
PageRank
Abhijit Guha Roy100.34
Jie Ren200.34
Shekoofeh Azizi300.34
Aaron Loh400.68
Vivek Natarajan500.68
Mustafa Ayoob Basil653.15
Nick Pawlowski700.34
Jan Freyberg800.68
Yuan Liu900.34
Zach Beaver1000.34
Nam Vo1100.34
Peggy Bui1200.34
Samantha Winter1351.43
Patricia MacWilliams1400.34
Greg S. Corrado1500.34
Umesh Telang1600.34
Yun Liu1700.34
Taylan Cemgil1801.35
Alan Karthikesalingam1971.48
Balaji Lakshminarayanan2027021.07
Jim Winkens2100.34