OSLTBDNet: Orthogonal softmax layer-based tuberculosis detection network with small dataset

Abstract

Tuberculosis, often known as TB, is a chronic bacterial infection, which damages the lungs. TB is one of the top 10 main causes of mortality in the world. It is essential to make an accurate and prompt diagnosis of tuberculosis (TB). In this work, a novel Orthogonal Softmax Layer-based Tuberculosis Detection Convolutional Neural Network (OSLTBDNet) is developed by leveraging the merits of depthwise separable convolution, tunable hyperparameters, inverted residual bottleneck block, and orthogonal softmax layer (OSL)-based classification. OSL maintains the orthogonality among weight vectors to boost class discrimination capability; thus improving classification results. It reduces coadaptation between parameters by discarding several connections from the fully connected convolution layer, FCCL, hence simplify the optimization. Thus, leveraging these above-discussed salient features makes the proposed system more accurate and faster as well. The results of this experiment suggest that the proposed model is superior to other comparing models with the best 99.00%, and 98.17% accuracy in Kaggle and TBX11K datasets, respectively.