Deep learning classification of active tuberculosis lung zones wise manifestations using chest X‐rays: a multi-label approach
James Devasia, Hridayanand Goswami, Subitha Lakshminarayanan, Manju Rajaram, Subathra AdithanAim: To develop a model that performs multi-label classification of manifestations of active tuberculosis in chest-xrays using EfficientNetB4 architecture and evaluate its performance
Methods: The total dataset comprised 1312 chest x-ray (837 patients), divided into 262(20%) for testing and 1050(80%) for training/validation (840-training, 210 validation). Offline augmentation of 840 training images was done to produce 2520 images. Annotation of the training and validation data set was done using 19 labels. On tabulation of annotations, it was found that only 12 labels had more than 10% frequency in the dataset and were taken as the primary final classifiers. An additional 30 classifiers were generated, based on the six lung-zonal ( right/ left upper/mid/lower zones) locations of five of these 5 labels - opacity, cavity, fibrosis, calcification, and collapse). A further additional 14 classifiers were made from the right/left location of 7 labels- pleural thickening, pleural effusion,pneumo/hydrothorax, tracheal shift, mediastinal shift, volume loss, emphysema. Thus the secondary final classifier had 44 labels. These xrays were trained using EfficientnetB4 (pre-trained with ImageNet). ImageNet’s 1000 classes were replaced with 4 layers, namely, 1. Global Average Pooling layer (GAP) 2. a Batch Normalisation layer ( 4 mini matches) 3. Dropout layer 4. a classifier layer with 12/44 output classes and Sigmoid activation. Progressive training was done to obtain the model. The model was evaluated on testing and external dataset (Montgomery-County-CXR-Set)
Results: The model performed well in the testing and external dataset. The opacity label had the highest accuracy, precision, and F1 score.
Conclusion: Multi-label classifications of manifestations of active TB can be used to develop a model to diagnose TB. The model performed with high accuracy, precision, and F1 score, indicating a potential to deploy in real-world primary care settings.