Abstract
Aim: Tuberculosis is an infectious disease caused by Mycobacterium tuberculosis bacteria. This study plans to build a novel deep learning-based model for the accurate recognition of tuberculosis. Methods: We propose a novel model - rotation angle vector grid-based fractional Fourier entropy and deep stacked sparse autoencoder (RAVG-FrFE-DSSAE) - which uses RAVG-FrFE as a feature extractor and harnesses DSSAE as the classifier. Moreover, an 18-way MDA is introduced on the training set to avoid overfitting. Results: Experimental results of 10 runs of 10-fold CV showcase that this proposed RAVG-FrFE-DSSAE algorithm yields a reasonable performance including of 93.68 +/- 1.11% sensitivity, 94.38 +/- 1.11% specificity, 94.35 +/- 1.04% precision, 94.03 +/- 0.69% accuracy, 94.01 +/- 0.70% F1-score, 88.07 +/- 1.38% MCC, 94.01 +/- 0.70% FMI, and 0.9725 AUC, respectively. Conclusions: Our result outperforms the eight state-of-the-art approaches. Besides, the result shows the effectiveness of the 18-way MDA.