Optimized Residual U-Net with Ant Colony Hyperparameter Tuning for Lung-Field Segmentation in Chest X-Ray Images

Abstract

To address the challenges of hyperparameter sensitivity and loss of fine details in medical image segmentation, particularly in lung-field segmentation from chest X-ray images, this study introduces a hybrid neural network architecture named UNet-ResNet50-ACO. The proposed framework integrates the Ant Colony Optimization (ACO) algorithm with a deep residual U-Net. The ACO component establishes a dynamic parameter optimization mechanism through efficient bio-inspired global search, enabling automated hyperparameter tuning. Simultaneously, residual skip connections within the encoder enhance feature propagation and mitigate network degradation, thereby improving the preservation of fine lung-boundary and anatomical features. Evaluated on the COVID-19 Radiography Dataset, the method achieves a Dice coefficient of 97.36%, representing an improvement of approximately 18 percentage points over the standard U-Net baseline, along with concurrent gains in IoU. Ablation studies confirm the individual contributions of both the ResNet50 backbone and the ACO algorithm. Moreover, the model demonstrates robust lung-boundary delineation under low-contrast chest X-ray conditions. The main contributions of this work are: (1) a collaborative ACO-CNN mechanism for intelligent hyperparameter space exploration; (2) suppression of semantic attenuation through the integration of residual learning and the U-Net architecture; and (3) a modular and highly adaptable framework for multimodal medical image analysis, offering an effective preprocessing framework for downstream pulmonary image analysis and computer-aided diagnosis. In addition, comparisons with representative CNN-based, Transformer-based, and foundation-model-based segmentation methods under the same experimental protocol demonstrate the competitive performance of the proposed framework.