P344 Automatic detection of pneumocystis jirovecii in microscope images: adeep learning-based approach

: Poster session 3, September 23, 2022, 12:30 PM - 1:30 PM OBJECTIVE: Pneumocystis jirovecii Pneumonia is one of the diseases that most affect immunocompromised patients today, and under certain circumstances, it can be fatal. One of the most widely used techniques in diagnostic laboratories for the detection of its etiological agent is optical microscopy. However, some of the disadvantages of this technique are its low sensitivity, low accuracy, and high dependence on an expert to make the diagnosis. Then, this work aims to develop a computational tool based on a deep learning approach to automatically detect the presence of P. jirovecii Pneumonia fungus from optical images, and to increase the accuracy of this conventional technique. METHODS: The study involved 29 randomized patients, from whom respiratory samples (bronchial lavage, and bronchoalveolar lavage) were collected. Methenamine silver staining was then used to prepare the samples. Subsequently, the slides of the analyzed patients were observed using the Leica DM500 microscope using a Leica ICC50 HD camera, and the optical images were taken in at least four random positions on the specimen holder. Thus, an image dataset of 29 different patients was created to detect whether a patient is positive or negative for P. jirovecii Pneumonia. Finally, a deep learning approach based on convolutional neuronal networks (CNN) was proposed and evaluated to improve the accuracy of the microscopy technique. The proposed CNN method incorporates global and local features for pixel-wise segmentation. RESULTS: First, the dataset image was processed and segmented using the connected components methodology. Likewise, the segmented images were labeled with the help of an expert to train the algorithm. To validate the response of the proposed deep learning approach the obtained results were compared with the obtained conventional image classification techniques like co-occurrence matrix and K-NN. The obtained results reveal that the proposed methodology allows to increase in the accuracy in the P. jirovecii Pneumonia identification up to 98%, while the co-occurrence matrix and K-NN only achieve accuracies of 89% and 85% respectively. CONCLUSION: It is possible to demonstrate that techniques based on digital image processing are a useful tool to support the processes of analysis and diagnosis of samples in medical patients with P. jirovecii Pneumonia. In addition, the obtained results demonstrate that methods based on deep learning allow us to develop more precise and accurate analysis methodologies for the analysis of patient samples with P. jirovecii Pneumonia. Our model can be improved by adding new layers, but this would introduce even more hyperparameters that should be adjusted. We intend to extend our model architecture in other areas of medical imaging with the usage of deep learning and computer vision techniques.