Keywords: Learnable test-time augmentation, domain shift, medical image segmentation
TL;DR: A novel learnable test-time augmentation tailored specifically to alleviate large domain shifts for the source-free medical image segmentation task
Abstract: As distribution shifts are inescapable in realistic clinical scenarios due to inconsistencies in imaging protocols, scanner vendors, and across different centers, well-trained deep models incur a domain generalization problem in unseen environments. Despite a myriad of model generalization techniques to circumvent this issue, their broad applicability is impeded as (i) source training data may not be accessible after deployment due to privacy regulations, (ii) the availability of adequate test domain samples is often impractical, and (iii) such model generalization methods are not well-calibrated, often making unreliable overconfident predictions. This paper proposes a novel learnable test-time augmentation, namely OptTTA, tailored specifically to alleviate large domain shifts for the source-free medical image segmentation task. OptTTA enables efficiently generating augmented views of test input, resembling the style of private source images and bridging a domain gap between training and test data. Our proposed method explores optimal learnable test-time augmentation sub-policies that provide lower predictive entropy and match the feature statistics stored in the BatchNorm layers of the pretrained source model without requiring access to training source samples. Thorough evaluation and ablation studies on challenging multi-center and multi-vendor MRI datasets of three anatomies have demonstrated the performance superiority of OptTTA over prior-arts test-time augmentation and model adaptation methods. Additionally, the generalization capabilities and effectiveness of OptTTA are evaluated in terms of aleatoric uncertainty and model calibration analyses. Our PyTorch code implementation is publicly available at https://github.com/devavratTomar/OptTTA.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: both
Primary Subject Area: Application: Other
Secondary Subject Area: Transfer Learning and Domain Adaptation
Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.
Code And Data: https://github.com/devavratTomar/OptTTA