Chest X-ray pneumothorax segmentation using U-Net with EfficientNet and ResNet architectures.

Area of Science:

• Medical imaging and computational analysis.
• Application of artificial intelligence in healthcare.
• Development of advanced deep learning models for image processing.

Background:

• Medical image segmentation is crucial for diagnosis but is often manual, time-consuming, and error-prone.
• Deep learning, particularly semantic segmentation networks like U-Net, offers automated pixel-level image understanding.
• Existing methods necessitate improvements in accuracy and efficiency for clinical applications.

Purpose of the Study:

• To introduce a novel, end-to-end semantic segmentation model for medical images named Ens4B-UNet.
• To improve the accuracy and efficiency of medical image segmentation through an ensemble approach.
• To address the limitations of manual segmentation in clinical settings.

Main Methods:

• Developed Ens4B-UNet, an ensemble model combining four U-Net architectures with pre-trained convolutional neural network (CNN) backbones.
• Employed nearest-neighbor up-sampling in decoders and leveraged transfer learning from ImageNet.
• Utilized techniques including stochastic weight averaging (SWA), data augmentation, test-time augmentation (TTA), and optimal thresholding.

Main Results:

• Achieved a mean Dice Similarity Coefficient (DSC) of 0.8608 on the 2019 Pneumothorax Challenge dataset.
• The model demonstrated performance within the top one percent of systems in the Kaggle competition.
• The ensemble approach with pre-trained backbones and advanced training techniques proved effective.

Conclusions:

• Ens4B-UNet represents a significant advancement in automated medical image segmentation.
• The proposed model offers a robust and accurate solution for analyzing medical images, aiding in clinical diagnosis.
• The findings highlight the potential of ensemble deep learning methods in medical image analysis.