Diagnóstico da tuberculose em microscopia de campo claro usando redes profundas

Abstract

The use of automatic smear microscopy for the diagnosis of pulmonary tuberculosis has been the subject of many studies published in recent decades. Most of them deal with a preliminary stage of diagnosis, the bacilli detection, while, as determined by World Health Organization (WHO), smear microscopy comprises the detection and reporting of the number of bacilli found in up to 100 microscopic fields for the prescription of the diagnosis that can be Negative, Scanty, 1+, 2+ or 3+. The diagnosis of pulmonary tuberculosis using bright-field smear microscopy depends on the attention of a trained technician who, with high laboratory demands, can be overloaded. While automated diagnosis requires little or no interpretation by a technician. As far as we know, this work proposes the first automatic method for diagnosing pulmonary tuberculosis using bright-field smear microscopy, according to the WHO guidelines. The proposed method comprises a semantic segmentation step, using a deep neural network, followed by a filtering step aiming to reduce the number of objects detected that are not bacilli (false positives): filtering based on color and shape features of the object. In semantic segmentation step, different encoder configurations were evaluated, using depth-wise separable convolution layers and channel attention mechanism. In the filtering step, two models were evaluated, the first with a color filter and a shape filter, and the second with two color filters and a shape filter. The second filtering model showed better performance. The proposed method was evaluated with a large, robust, and annotated dataset designed for this purpose, consisting of 250 test sets, where each diagnostic class has 50 sets. The following performance metrics were obtained with the proposed method for classifying the 5 diagnostic classes: average precision of 0.894, average recall of 0.896 and average f1-score of 0.895. Furthermore, the method presented a diagnostic time of approximately 7 minutes in the case where more digital fields are needed for the analysis. Therefore, the results achieved showed the possibility of automatically diagnosing tuberculosis using bright-field smear microscopy.