Title:An Optimal Model Combining SqueezeNet and Machine Learning Methods for
Lung Disease Diagnosis
Volume: 20
Author(s): Abdallah Maiti*, Abdallah Abarda, Mohamed Hanini and Ahmed Oussous
Affiliation:
- Laboratory of Computing, Networks, Mobility and Modelling (IR2M) FST, Hassan First University of Settat, Settat, Morocco
Keywords:
Decision tree, Deep learning, KNN, SVM, Logistic regression, SqueezeNet, Lung diseases diagnosis.
Abstract:
Background:
Artificial intelligence (AI) is rapidly evolving in healthcare, with transformative potential. AI revolutionizes medical imaging by enabling online
self-diagnosis for patients and improving diagnostic accuracy for healthcare professionals. While valuable datasets aid machine learning in disease
detection, challenges persist in diagnosing similar lung conditions from chest X-rays. Integrating AI into healthcare holds promise for enhanced
outcomes and efficiency.
Objective:
In this article, we aim to present a new AI model that solves this challenge by allowing the differentiation, diagnosis and classification of three
distinct diseases, whose symptoms are very similar. The fundamental contribution is to reduce the number of parameters used while maintaining
the same level of precision for use in embedded systems.
Methods:
Our proposed model combines the power of the neural network using the SqueezeNet architecture with a set of machine learning algorithms as
classifiers, including logistic regression, support vector machine (SVM), k-nearest neighbors (KNN), decision tree, and naive Bayes. The chest Xray
dataset used in the proposed model consists of CXR images that are classified into four categories: pneumonia, tuberculosis, COVID-19, and
normal cases.
Results:
Our proposed model demonstrated remarkable accuracy (97,32%), precision (97,33), F1 score (97,31%), recall (97,30%), and AUC (99,40), which
is close to the best model. Whereas, the number of parameters used by our model (4,6 M) is very small compared to the best model in the literature
(47M).
Conclusion:
The model demonstrated good classification accuracy. In addition, the proposed model has the ability to use fewer parameters, which means it
requires less internal memory and computing resources.
