Universal Model for 3D Medical Image Analysis

Deep Learning-based methods recently have achieved remarkable progress in medical image analysis, but heavily rely on massive amounts of labeled training data. Transfer learning from pre-trained models has been proposed as a standard pipeline on medical image analysis to address this bottleneck. Despite their success, the existing pre-trained models are mostly not tuned for multi-modal multi-task generalization in medical domains. Specifically, their training data are either from non-medical domain or in single modality, failing to attend to the problem of performance degradation with cross-modal transfer. Furthermore, there is no effort to explicitly extract multi-level features required by a variety of downstream tasks. To overcome these limitations, we propose Universal Model, a transferable and generalizable pre-trained model for 3D medical image analysis. A unified self-supervised learning scheme is leveraged to learn representations from multiple unlabeled source datasets with different modalities and distinctive scan regions. A modality invariant adversarial learning module is further introduced to improve the cross-modal generalization. To fit a wide range of tasks, a simple yet effective scale classifier is incorporated to capture multi-level visual representations. To validate the effectiveness of the Universal Model, we perform extensive experimental analysis on five target tasks, covering multiple imaging modalities, distinctive scan regions, and different analysis tasks. Compared with both public 3D pre-trained models and newly investigated 3D self-supervised learning methods, Universal Model demonstrates superior generalizability, manifested by its higher performance, stronger robustness and faster convergence. The pre-trained Universal Model is available at: \href{https://github.com/xm-cmic/Universal-Model}{https://github.com/xm-cmic/Universal-Model}.