UNet++ and Efficient Net: Hyperparameter Optimization Training Strategy for Precise Pneumothorax Image Segmentation

The lungs play a crucial role in the respiratory analysis of human beings. Pneumothorax is a condition that arises from the leakage of air into the pleural space, resulting in the collapse of the lungs. This condition can impede respiratory movement, cause breathing difficulties, disrupt blood flow, lower blood pressure, lead to pleurisy and cardiovascular disorders, and in severe cases, pose a life-threatening risk. Therefore, the detection and timely treatment of pneumothorax are imperative. While pneumothorax is typically identified through the primary diagnostic imaging technique of chest X-ray (CXR) imaging, the effectiveness of treatment relies on a prompt review of radiographs. A computerized diagnosis system has the capability to identify pneumothorax in chest radiographic images, offering significant advantages in disease diagnosis. In the current study, a novel architecture comprising UNet++ and the pre-trained EfficientNet V4 model is proposed for the detection of pneumothorax regions in chest X-ray images. The novelty of the proposed method resides in its effective utilization of the hyperparameter optimization strategy. This approach involves training models with a diverse range of architectures and training parameters, enhancing the efficiency of the overall methodology. It helps in fine-tuning the model to achieve the best possible performance, ultimately contributing to more reliable and accurate segmentation results. Also, BCE dice loss is primarily considered as the main criterion to assess the model’s accuracy. The higher the BCE dice loss better the model and the output segmentation. This study gives the highest BCE dice loss of 0.89, and dice loss for validation images is 0.86. In order to achieve this value of BCE dice loss, we did a few iterations to zero in on the particular learning rate. Number of iterations have been carried out with different preprocessing of image, batch size, and learning rate to obtain the highest possible BCE dice loss. Thus, the NestedUNet++ architecture offers a promising solution for the accurate segmentation of pneumothorax images. By leveraging the power of deep learning and advanced architectural features, NestedUNet++ with hyperparameter optimization demonstrates its potential to aid in the diagnosis and treatment of pneumothorax, contributing to improved healthcare outcomes and patient care.