Vision Mamba Network For Tuberculosis Detection From Chest X-Rays

Detecting Tuberculosis (TB) in chest radiographs is not a simple problem for a model because it has to notice very small, localized lesions while also keeping the overall context of both lung fields, and the aim here is to address this local-plus-global challenge.In this paper, MSVMamba-TB is a hierarchically nested multi-scale Vision Mamba network classifier of 1 channel chest X-ray (CXR) into normal and TB defined in binary classification.We compared it with a deep learning technique, namely Vision Transformer (PaperViT), and two other approaches: ConvNeXtV2-Pico and EfficientNet-B3.Not only that, but all the models were also used in conjunction, fine-tuned, and verified on the same computer with the same specs and setup to keep the comparison for every model as similar as possible.In our proposed model, a multi-scale two-dimensional selective scan, a convolutional feed-forward block, squeeze-andexcitation channel weighting, global pooling, and a small binary head are put together in one architecture.Experiment A contained a total of 4, 200 patient-level records, of which 3, 500 were normal and 700 were tuberculosis.The data distribution was an 80/10/10 split, whereas Experiment B consisted of 7, 000 balanced records with a 70/15/15 data split.Both scenarios have the same preprocessing, augmentation, sampling, loss, optimizer, scheduler, early stopping, and other model features, and no model is given a separate advantage in terms of training.In Experiment A the model obtained accuracy of 99.05%, a sensitivity of 97.14%, and precision of 97.14%, with the F1-score of 97.14% being the experimental results.Experiment B was able to reach an accuracy of 98.38%, a precision of 99.42%, a sensitivity of 97.33%, an F1-score of 98.36%, and the highest score was on the AUC at 0.9992.The model proposed here ranked top in both accuracy and F1-score and made fewer mistakes.Therefore, our proposed model aims to have good prediction performance and a small size.Other than that, this is a very small model with only 7.012M trainable parameters, which, relative to PaperViT with 13.527 M, is only 48.17%.Detailed epoch records, confusion matrices, ROC curves, error analysis, activation maps, and cases where failures were made are presented for all four networks.Hence, the results are not only reported as a single accuracy number.Moreover, these results indicate that MSVMamba-TB is an encouraging candidate for fast tuberculosis screening, which is useful for computer-aided TB screening.