Abstract:
Tuberculosis is a radiological syndrome that can be fatal if not treated properly in time. This project pertains to usage of neural networks for the fully automated diagnoses of tuberculosis from chest x-rays. The proposed system works by first enhancing the contrast of the candidate scan. Then, it is digitalized by keeping 80% of the highest intensity pixels. Afterwards, the torso mask is generated by iteratively analyzing the binaries x-rays scan. The torso mask is then XORed with the binaries scan to extract lungs mask. The segmented mask is then post processed to remove noisy outliers and it is then multiplied with the original chest x-rays image to extract lungs. The segmented lungs scan is then passed to the pertained Alex Net convolution neural network model for tuberculosis diagnosis. The proposed system has been tested on two publicly available dataset and to the best of our knowledge it is one of the generalized frameworks that diagnosis tuberculosis irrespective of the scan quality and scan acquisition machinery. Proposed system achieved the diagnostic accuracy of 95.01% on both datasets.