AFA: adversarial frequency alignment for domain generalized lung nodule detection.

The different reconstruction parameters of CT imaging lead to domain shifts, which limits the generalization of deep learning models and their applications in computer-aided diagnosis systems. In this paper, we investigate the multi-source domain generalization (DG) problem in the context of lung nodule detection from CT images. We first identify the reconstructed convolution kernel as the key parameter leading to domain shifts. Accordingly, we reorganize the public LUNA16 dataset into a domain generalization benchmark, i.e.,, LUNA-DG. Then, we propose a novel DG method by adversarial frequency alignment (AFA). Specifically, we devise an adaptive transition module (ATM) to learn a frequency attention map that can align different domain images in a common frequency domain. For this purpose, a fidelity discriminator and a multi-domain discriminator are used to train the ATM alternately and adversarially. In addition, to mitigate the issue of ineffective gradient back-propagation in naive multi-domain adversarial learning, we propose a novel random domain adversarial learning (RDAL) strategy that can back-propagate effective gradient signals and gradually reduce the gap between multiple domains. The ATM can be combined with nodule detection models through differentiable Fast Fourier Transform (FFT) and inverse FFT, allowing end-to-end training. Experimental results on both LUNA-DG and our in-house datasets validate the superiority of AFA over representative DG methods. [ABSTRACT FROM AUTHOR]