Objective. To estimate the possibility of diagnosis of malignant pleural effusion using convolutional neural networks of facies images of pleural exudates obtained by the method of wedge-shaped dehydration. Material and methods. We studied 163 images of pleural fluid facies obtained by wedge-shaped dehydration in patients with various pleural effusions (10 nosological groups). Recognition and analysis were carried out using convolutional neural network. The images were divided into two groups — malignant effusion (n=65; 40%) and other diseases (n=98; 60%). Results. There were 131 photos selected for further investigation after pre-processing of images by eliminating defective ones, turning them into black and white format, cleaning of ‘noise’, cutting out the facies. Then the images were standardized. The method of rigid transformations with rotation for every 10 degrees was used. As a result, their number increased up to 4,585. Self-taught neural network analyzed the images of facies independently by separation of the fragments consisting of black and white dots and comparison of them with each other. Self-teaching and training of each neural network were ensured by random sampling of 80% of images from the initial sample. Then the remaining 20% of the images were used as a control sample to assess the possibilities of recognition pleural effusion cause. Four options of convolutional neural networks were used. An accuracy of cancer detection ranged from 82% to 95.6%, benign diseases — from 84% to 94.7%. The neural network with the highest sensitivity was chosen. Conclusion. Automated image analysis system of pleural effusion facies using convolutional neural network ensured an accuracy of diagnosis of malignant pleural effusion in 95,6% of cases and other diseases in 90% of cases. The method is simple, efficient, cheap and reagentless.