Enriching Chest Radiography Representations: Self-Supervised Learning With a Recalibrating and Importance Scaling

Chest radiography (CXR) is a widely researched area in medical imaging processing due to its diagnostic utility in heart-related and thoracic diseases. However, the scarcity of labeled data poses challenges for fully supervised approaches. Self-supervised learning (SSL), capable of deriving meaningful representations from unlabeled data, offers a promising solution. Unfortunately, most existing self-supervised instance discrimination methods primarily focus on learning global invariant representations, whereas local spatial representations are pivotal in diagnosing diseases within the CXR domain. Moreover, due to substantial differences between natural images and CXR images, such as color distribution and texture, the performance of SSL methods with CXR data remains uncertain. To address these challenges, we propose a novel unit called the Recalibrating and Importance Scaling Layer (RS-Layer), which aims to learn adequate representation from CXR data to provide more fine-grained and discriminative features for diverse downstream tasks. The RS-Layer consists of a recalibrating module that extracts general features from compressed features and an importance scaling module that assesses the significance of each feature. By emphasizing crucial feature components based on their calculated importance and suppressing non-essential features, the RS-Layer empowers the SSL method to capture valuable information for meaningful representations explicitly. We conduct a systematic analysis to evaluate the effectiveness of the RS-Layer from three key perspectives: 1) Demonstrating its generalizability and transferability across diverse downstream tasks, 2) Analyzing the feature space it produces, and 3) Conducting an extensive ablation study on the components that constitute the RS-Layer. Importantly, the RS-Layer offers the advantage of being adaptable to existing SSL instance discrimination methods