[SCI(E)-IEEE Access] Enriching Chest Radiography Representations: Self-Supervised Learning with a Recalibrating and Importance Scaling - Heesan Kong, Donghee Kim, Kwangsu Kim
- 인공지능 융합 연구실
- 조회수228
- 2023-09-24
Abstract
Chest radiography (CXR) is the most widely investigated field in medical imaging processing owing to its usefulness in diagnosing heart-related or other thoracic diseases. However, due to the scarcity of labeled data, there is a constraint on using a fully supervised approach. Self-supervised learning (SSL), which can acquire meaningful representations from unlabeled data is an effective method to solve this problem. Unfortunately, most of the existing self-supervised instance discrimination methods aim to learn global invariant representations, whereas local spatial representations for disease diagnosis are crucial for the CXR domain. Furthermore, because there are noteworthy differences between natural and CXR images, such as color distribution and texture, it is not clear how well the SSL method performs for CXR data. To alleviate these issues, 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 a more fine-grained and discriminative features for various downstream tasks. The RS-Layer consists of a recalibrating module that extracts general features from squeezed features and an importance scaling module that determines the importance of each feature. By emphasizing the crucial parts of a feature according to its calculated importance and suppressing non-essential features, the RS-Layer enables the SSL method to explicitly capture the useful parts for learning valuable representations. A systematic analysis was conducted to investigate the efficacy of the RS-Layer from three perspectives: 1) Proof of generalizability and transferability for diverse downstream tasks, 2) Analysis of the feature space, and 3) Extensive ablation study of the components constituting the RS-Layer. In addition, the RS-Layer has the advantage of being applicable to existing SSL instance discrimination methods.
- 다음글
- 다음글이 없습니다.