Multi-label Classification Technique of Chest X-Rays Image Based Cardiomegaly Disease Prediction

One of the most common techniques used in detecting serious life-threatening diseases is chest X-Ray radiography, through which datasets of X-ray images are collected. Among the serious heart-related chest diseases that can be detected using this technique is Cardiomegaly. However, the X-ray image-based identification of diseases considers a time-consuming process and requires radiologists with high skills to interpret and analyze these images to accurately diagnose the pathologies, especially for difficult cases that cannot be interpreted by the naked eyes of humans where one image may have more than one pathology. In this context, to solve the above problems. in this paper, we deal with the above problems by designing the efficient architecture of two automated classification models based on the transfer learning technique and using the convolution neural network DenseNet121 as feature engineering. These architectures were constructed from DenseNet121 as a backbone model followed by a proposed deep neural network which consists of a global average pooling layer, two neural network layers, and an output layer. The first model was designed as a multi-label classification model to predict 8 types of chest diseases, thus the output layer contains 8 neurons and a sigmoid activation function, while the second model focused on the binary classification of cardiomyopathy so the output layer contains one neuron. Two custom functions were proposed for multi label classification model to be suitable for its task, loss function, and calculation of accuracy function. The performance of the models was implemented on the CheXpert dataset and evaluated in terms of accuracy and area under the curve (AUC). The results show that the first model achieved an AUC score of 90% and the second model obtained 83% which consider promising results. In addition, a web application interface was produced in this work which contributed to the practicality applicable to the model where it was examined in practical clinical to prove the generalization performance of the models, where the testing results were good and realistic.