Automatic Detection of Tuberculosis Bacilli in Sputum Smear Scans Based on Subgraph Classification

Tuberculosis (TB) is a chronic respiratory disease with high infectivity and mortality. Early diagnosis is important for curing TB and epidemic prevention. Clinically, sputum smear microscopy examination is a widely used method for TB examination. But it requires doctors to detect and count TB bacilli manually, which is laborious and error prone. Even though many semi-automatic or automatic methods have been proposed to detect TB bacilli, there are still some problems: a) Sputum smear microscopic images are shot by choosing field of view manually, b) Images have low resolution, c) Labeling TB bacilli is a huge workload. In our experiment, we adopted sputum smears images scanned by the high-resolution slide scanning system. Considering the characteristics of the images, we proposed a dataset construction strategy based on non-overlapping subgraph partition. To evaluate this method, we used three well-known convolutional neural network models (Inception v3, ResNet, DenseNet) on a dataset of 2,630 sputum smear microscopic images. The experiment results got best performances on Inception v3 with all indicators were above 98%. Then we stitched predicted results of subgraphs for display. The results reached the WHO criteria that sputum slide reading diagnosis error rate should less than 5%. This method can provide doctors with a wider and visualized view to identify TB bacilli in sputum smear scans, which means improvement of the diagnosis efficiency.