P0091DEEP LEARNING-BASED IMMUNOFLUORESCENCE ASSESSMENT OF GLOMERULAR DISEASES

Background and Aims Immunofluorescence (IF) tests of renal tissue are of great value in diagnosing most of the glomerular diseases. However, high quality IF tests and results interpretation by experienced pathologists are not universally available in different areas of China. The development of deep neural networks has been used to facilitate digital analysis of pathologic images recently. We proposed a novel Convolutional Residual Dense Network (CR-DenseNet) to facilitate IF assessment of renal biopsy samples. Method A dataset with 725 IF images, including 312 images of IgA nephropathy (IgAN), 319 images of idiopathic membranous nephropathy (IMN) and 94 images of type V lupus nephritis (LN V) diagnosed in Peking Union Medical College Hospital (PUMCH) from November, 2016 to March, 2018 were used for training and validation of CR-DenseNet. All the images were photographed using ANDOR, ZYLA, Japan. The resolution was 2560 × 2160 pixels. These images were carefully annotated for the distribution characteristics and final diagnosis by two renal pathologists independently. IgAN showed mesangial predominating deposition. IMN showed glomerular basement membrane (GBM) predominating deposition. LN V showed positive depositions in both mesangial area and GBM. These three groups were used for IF distribution identification training. In CR-DenseNet, convolutional residual dense blocks were introduced, each of them consisted of a dense block with a convolutional skip connection to fully exploit the dense local features. To identify the deposition location, we imposed a switch in the proposed model to handle an additional input for different types of tasks, which can provide glomerular contours approximated from the IF image foreground (Figure 1). Performance was evaluated using overall accuracy and F1 score. F1 was computed as 2×True Positive / (2×True Positive + False Positive + False Negative). The whole protocol was approved by Institutional Review Board of PUMCH (No. S-K913). Results Experimental results showed that the proposed CR-DenseNet model outperformed the state-of-the-art method. In identification of segmental and glomerular IF positive images, our model showed overall accuracy of 92.0%. The true positive rate and F1 score of segmental positive samples recognition were 90.1% and 0.889. The true positive rate and F1 score of glomerular recognition positive samples were 94.1% and 0.938. Furthermore, the overall accuracy of our model in identifying mesangial predominating deposition (IgAN), GBM predominating deposition (IMN) as well as positive depositions in both mesangial area and GBM (LN V) was 91.2%. The true positive rates of each above-mentioned deposition classification were 93.1%, 90.5% and 88.9%, respectively. The corresponding F1 scores were 0.964, 0.905 and 0.889. Conclusion Our preliminary data showed that CR-DenseNet model was quite powerful in making IF diagnosis of typical glomerular diseases for the first time.