Unsupervised Cross-domain Pulmonary Nodule Detection without Source Data

TL;DR

This work proposes a Source-free Unsupervised cross-domain method for Pulmonary nodule detection (SUP), named Instance-level Contrastive Instruction fine-tuning framework (ICI), which first adapts the source model to the target domain by utilizing instance-level contrastive learning, and substantially surpasses existing state-of-the-art approaches.

Abstract

Cross-domain pulmonary nodule detection suffers from performance degradation due to a large shift of data distributions between the source and target domain. Besides, considering the high cost of medical data annotation, it is often assumed that the target images are unlabeled. Existing approaches have made much progress for this unsupervised domain adaptation setting. However, this setting is still rarely plausible in medical applications since the source medical data are often not accessible due to privacy concerns. This motivates us to propose a Source-free Unsupervised cross-domain method for Pulmonary nodule detection (SUP), named Instance-level Contrastive Instruction fine-tuning framework (ICI). It first adapts the source model to the target domain by utilizing instance-level contrastive learning. Then the adapted model is trained in a teacher-student interaction manner, and a weighted entropy loss is incorporated to further improve the accuracy. We establish a benchmark by adapting a pre-trained source model to three popular datasets for pulmonary nodule detection. To the best of our knowledge, this represents the first exploration of source-free unsupervised domain adaptation in medical image object detection. Our extensive evaluations reveal that SUP-ICI substantially surpasses existing state-of-the-art approaches, achieving FROC score improvements ranging from 8.98% to 16.05%. This breakthrough not only sets a new precedent for domain adaptation techniques in medical imaging but also significantly advances the field toward overcoming challenges posed by data privacy and availability. Code: https://github.com/Ruixxxx/SFUDA.

Figures (4)

• Figure 1: (a) Existing SFUDA object detection works utilize feature alignment or sample generation to help with the pseudo labeling. These approaches mainly focus on exploiting the source model. (b) Our proposed SUP-ICI utilizes instance-level contrastive learning (CL) to make use of the foreground-background semantic information of the unlabeled target images. Our weighted entropy (WE) loss is also incorporated for label denoising.
• Figure 2: Overview of our Source-free Unsupervised cross-domain method for Pulmonary nodule detection (SUP) named Instance-level Contrastive Instruction fine-tuning framework (ICI). The SUP-ICI consists of two steps: (a) initial pseudo nodule generation and (b) pseudo supervised pulmonary nodule detection. (a) The well-trained source model is adapted to the target domain by instance-level contrastive learning. The proposals of RPN are identified as nodules and non-nodule instances, and cropped from the feature maps as regions of interest (RoI). A positive pair is formed using two nodule or two non-nodule RoI features, whereas a negative pair is formed using one nodule and one non-nodule RoI features. Through contrastive learning, the positive pairs are pulled together and the negative pairs are pushed apart. (b) The adapted model is duplicated into a teacher model and a student model. The teacher is gradually updated by the student using EMA strategy to generate pseudo nodules for the student training. Furthermore, an unsupervised weighted entropy (WE) constraint is added to the R-CNN iccv15/fastrcnnnips15/fasterrcnn to alleviate the negative effect of the pseudo label noise. The WE loss can not only increase the prediction confidence, but also make the localization of the nodule more accurate.
• Figure 3: Samples of four pulmonary nodule datasets. Each column belongs to one pulmonary nodule dataset as labeled. CT images of different datasets present domain discrepancy, such as illumination, color contrast/saturation, resolution, number of nodules.
• Figure 4: Exemplar detection results of the compared approaches and our method SUP-ICI. The green boxes, red boxes, and yellow boxes denote the true positives (TP), false positives (FP), and false negatives (FN), respectively. The values above the boxes are the detection scores.