Optimizing the U-Net Model for Segmenting the Lung Opacity Regions in Chest Radiographs

The successful training of deep networks requires an enormous amount of annotated training samples which often necessitates powerful hardware and extended processing times leading to increased costs and potential delays. The primary purpose of convolutional networks is to serve classification tasks that assign a single label to an entire input. However, segmentation serves as a powerful solution to overcome the limitations inherent in classification by providing a more nuanced and detailed analysis of input data. Semantic segmentation contributes to a more granular and contextually rich interpretation of visual data by providing a pixel-level classification. This work involves the U-Net model to effectively capture the lung opacity features from the X-ray images. Lung opacity is a non-specific term that highlights an abnormal appearance in the lung region which requires timely diagnosis and appropriate treatment. To ensure continuous neuron activity during the training process, the ReLU activation in the U-Net model is replaced with the Leaky ReLU activation. The optimized U-Net model is trained and validated using 90 labeled lung opacity chest radiographs. U-Net's adaptability to learn infections of varied sizes along with its capacity to handle limited labeled data through data augmentation enhanced its performance in highlighting the lung opacity regions. The model is able to delineate and match the ground truth regions of interest in the X-rays with a dice coefficient of 91.02% and IoU (Intersection Over Union) of 84.15%, allowing healthcare professionals to accurately identify and focus on specific areas of concern in X-rays.