A Fully Connected Reproducible SE-UResNet for Multiorgan Chest Radiographs Segmentation

Deep learning (DL) models are a popular choice for resolving intricate issues in medical imaging, such as the classification of diseases, detection of anomalies, and segmentation of tissues in real-world scenarios. To be useful in these contexts, the models must be able to provide accurate results for new, previously untrained data. Existing methods not only fail to consider the intrinsic features of small target lesions but are also not evaluated on separate datasets. To solve these problems we propose a novel architecture, SE-UResNet, capable of segmenting multiple organs having different size and shapes from Chest X-Ray (CXR) images. The proposed architecture introduces a residual module in between the encoding and decoding modules of an attention U-Net architecture for better feature representation of high-level features. The architecture also replaces the attention gates in the decoder module of attention U-Net with Squeeze and Excite (S&E) modules. SE-UResNet is experimented on benchmark CXR datasets such as NIH CXR for lungs, heart, trachea and collarbone segmentation as well as VinDr-RibCXR for ribs segmentation tasks with respect to other state-of-the-art segmentation models. The proposed model achieves an average DSC of 95.9%, 76.8%, 78.7%, 78.8%, and 86.0% for lungs, trachea, heart, collarbone and ribs segmentation for the aforementioned datasets. Furthermore, the proposed model has only been tested on two benchmark CXR datasets: Shenzen and JSRT to establish the reproducibility and robustness of the model. The performance of SE-UResNet on several benchmark CXR datasets demonstrates the model’s ability to generalize, making it a reliable baseline for medical image segmentation. Furthermore, it can also be used for assessing the reproducibility of DL models based on their performance on different datasets.