Frontiers in medicine · 2026
Integrating augmentation-aware manifold smoothing and momentum-adjusted loss for handling class imbalance in thoracic disease detection.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
H R Sneha, B Annappa
Аннотация
INTRODUCTION: Class imbalance is a fundamental challenge in medical image analysis, where certain disease categories occur far less frequently than others. This uneven data distribution often causes learning algorithms to favor common conditions while underperforming on rare but clinically significant cases. In chest X-ray analysis, the deployment of artificial intelligence is particularly hindered by the long-tailed distribution of thoracic diseases, as conventional deep learning models exhibit optimization bias toward majority classes, leading to reduced sensitivity for rare yet critical pathologies.
METHODS: To address this challenge, this paper presents Dynamic Adaptive Weighting with Hybrid Networks (DAWN-Net), a unified framework that synergizes data-level and algorithm-level interventions. Unlike conventional approaches that treat augmentation and re-weighting in isolation, DAWN-Net introduces a Hybrid Synergy mechanism. At the data level, we propose Augmentation-Aware Manifold Smoothing, which generates synthetic variations in the local tangent space of minority samples to densify their feature representation. Architecturally, the model employs a dual-stream design comprising a Hierarchical Feature Propagation Network (HFPN) to capture high-frequency local textural details, and a Semantic Context Modeling Network (SCMN) to enforce global anatomical consistency. These components are jointly optimized using a novel Momentum-Adjusted Gradient Harmonization loss, which dynamically recalibrates gradient contributions based on batch-wise class statistics and augmentation intensity.
RESULTS: Validation on three large-scale benchmarks-NIH ChestXray14, CheXpert, and PadChest-demonstrates that DAWN-Net consistently outperforms state-of-the-art baselines, particularly in the detection of rare diseases such as Hernia and Fibrosis.
DISCUSSION: By mitigating the optimization bias and improving sensitivity for rare yet critical pathologies, DAWN-Net overcomes the limitations of conventional deep learning models, thereby offering a more reliable solution for safety-critical radiological diagnosis.
METHODS: To address this challenge, this paper presents Dynamic Adaptive Weighting with Hybrid Networks (DAWN-Net), a unified framework that synergizes data-level and algorithm-level interventions. Unlike conventional approaches that treat augmentation and re-weighting in isolation, DAWN-Net introduces a Hybrid Synergy mechanism. At the data level, we propose Augmentation-Aware Manifold Smoothing, which generates synthetic variations in the local tangent space of minority samples to densify their feature representation. Architecturally, the model employs a dual-stream design comprising a Hierarchical Feature Propagation Network (HFPN) to capture high-frequency local textural details, and a Semantic Context Modeling Network (SCMN) to enforce global anatomical consistency. These components are jointly optimized using a novel Momentum-Adjusted Gradient Harmonization loss, which dynamically recalibrates gradient contributions based on batch-wise class statistics and augmentation intensity.
RESULTS: Validation on three large-scale benchmarks-NIH ChestXray14, CheXpert, and PadChest-demonstrates that DAWN-Net consistently outperforms state-of-the-art baselines, particularly in the detection of rare diseases such as Hernia and Fibrosis.
DISCUSSION: By mitigating the optimization bias and improving sensitivity for rare yet critical pathologies, DAWN-Net overcomes the limitations of conventional deep learning models, thereby offering a more reliable solution for safety-critical radiological diagnosis.