Ensemble Stack Architecture for Lungs Segmentation from X-ray Images

In healthcare, chest X-rays are an inexpensive medical imaging diagnostic tools. The lung images segmentation from chest X-rays (CXRs) is important for screening and diagnosing diseases. The lungs are opacified in many patients' CXRs, making it difficult to segment them. A segmentation algorithm based on U-Net is proposed in this paper to address this problem. The proposed architecture was developed using three pre-trained models: MobileNetV2, InceptionResNetV2, and EfficientNetB0. In this architecture, we designed a ensemble stacked framework which is based on the pre-trained models to improve segmentation performance. Compared with the conventional U-Net model, our method improves by 3.02% dice coefficient and 3.43% IoU experimenting on the three public lung segmentation datasets.