Document is current

Article History

Received: 5 October 2023

Accepted: 17 January 2024

First Online: 8 February 2024

Declarations

:

: The authors declare that they have no conflict of interest.

: A considerable amount of work has been carried out in disease detection from a Chest X-ray, but the main limitations remain in multiple disease classification, stability of the network and solving the class imbalance problem were not focused much. Many a times, the researcher’s goal is to detect one disease but detecting all 13 diseases from one input image has not been approached before this work. Recent advances in deep learning have shown many good performances in disease identification from chest X-rays. The implementation of deep learning in medical imaging for disease detection from chest x-rays can help the doctors in early diagnosis and prognosis. The implementation of high accuracy AI models for medical imaging is challenging and has a major role in this field. Understanding the need of today’s world, we proposed a novel CX-Ultra net (Chest X-ray Ultranet) to classify and identify thirteen thoracic lung diseases from chest X-rays. Our contribution can have significant benefits for humanity. Reaching people who cannot get benefits of the latest medical facilities in city outskirts and unreachable places becomes possible with advancements in computer technology. We have used a multiclass cross-entropy loss function on a compound scaling framework using EfficientNet as a baseline. We performed channel shuffling in various stages of the network creating reduction cells and more skip connections which resulted in a high performance of the model. Also, there is a chance of higher negative data than the positive data containing the disease for many images database. A deep neural net is already hefty when trained with images that add to the computational complexity coupled with massive data. This eliminates considering the negative data for training purposes and hence the whole procedure of feature extraction and training them to the model, making the neural net heavier.