Multiple thoracic diseases detection from X-rays using CX-Ultranet

Recent developments in deep learning have demonstrated impressive performance in accurately identifying individual diseases from chest X-rays (CXRs). However, multiple diseases, stability of the deep network, and class imbalance problems were not addressed with high accuracy for disease detection and classification. So, the main purpose of this work is to develop a fully automatic computer method to detect thirteen types of thoracic disease from CXRs with high accuracy.

In this research, a CX-Ultranet has been proposed for the classification and detection of 13 different thoracic disorders from plain radiographic images. The baseline model employed is EfficientNet, and a multiclass cross-entropy loss function is utilized within a compound scaling structure. Channel shuffling is implemented at various stages of the network, creating reduction cells and incorporating more skip connections. The loss function algorithm and Adam optimizers work synergistically to stabilize the model and facilitate continuous learning from new data over time.

The CX-Ultranet demonstrates an average prediction accuracy of 88% when applied to diverse CXR datasets. In comparison to existing state-of-the-art techniques, the CX-Ultranet exhibits a remarkable improvement ranging from 5% to 15%. Additionally, it shows a reduction in operational time by approximately 30% compared to current cutting-edge models under similar environmental and data conditions.

The proposed CX-Ultranet achieves superior overall accuracy and effectively addresses imbalanced classes within the dataset. Furthermore, it significantly reduces the duration of network training in relation to FLOPS, thereby establishing a novel benchmark in the field of CXR-based disease diagnosis.