Acceleration Glomeruli Image Disease Detection using Deep Transfer Learning
Oleh : Ahmad Fauzan Aqil, S.Pd., M.Sc.
During the current pandemic, many people are susceptible to disease by the spread of COVID-19 in our bodies. For people who have comorbid diseases, COVID-19 has a longer resistance in the body. Although the main target of COVID-19 is the respiratory organ, not a few who are infected with COVID-19 appear new diseases stimulated by this virus. One of the most frequently heard is the emergence of chronic kidney disease caused by the penetration of this virus. According to George et al. was found that the COVID-19 content derived from angiotensin-converting enzymes 2 (ACE2) was able to penetrate human body tissues to damage the glomerular nucleus in kidney cells. This enzyme breaks down the contents of Bowman’s capsule from the glomerulus so that it is reduced to other cell parts in the kidney tissue. With such a scheme, George et al. prove that 25% of COVID-19 cases result in chronic kidney disease.
Such conditions are a concern to develop research that can solve these problems. The characteristics of the kidney glomerulus are straightforward to segment are feasible to accelerate the detection of kidney glomerular disease in humans. The part of the glomerulus of the kidney consists of the glomerular nucleus covered by Bowman’s capsule and the cell membrane around the glomerular nucleus. This structure becomes the main focus in observing medical images when there is a change in the structure and shape of the composition of the renal glomerulus. Image processing on kidney images using a CT scan that is extracted into biomolecular images. This image is then processed for further observation using deep transfer learning.
Deep transfer learning contains initial learning based on image objects that have been processed in machine learning. The results of the process are then reused as a pre-trained model to record images similar to the previous processing object. In this case, deep transfer learning is used specifically to detect medical images in kidney glomeruli disease which is divided into three classification classes, namely sclerosed, focal-segmental and normal glomeruli. The difference in the structure of the three classifications is that the sclerosed glomeruli have a more refractive cell structure than normal glomeruli, while the focal-segmental features still have Bowman’s capsule circumference which almost fades in the glomerular nucleus. The knowledge that has been developed regarding the different types of kidney glomerular disease becomes a reference for accelerating the detection of kidney glomerular disease diagnosis.
The performance of deep transfer learning contains a dataset containing the results of training data from a collection of biomolecular images associated with various shapes and characters that have previously been classified in the segmentation process of kidney tissue. The results of this training are processed to identify the classification of glomerular disease automatically using a test dataset. Then, the results of this classification become the final result of the learning process assessment that meets one of the classification classes. The higher the accuracy value of the resulting classification results, the more reliable the prediction results of the resulting classification. Using this method, this study found a suitable model for the classification of glomerular disease with high speed and accuracy using ResNet101V2.
ResNet101V2 has a detection accuracy value of 98.16% with a reading speed of only up to 5 seconds after using the loaded training database. This result is proven by experimenting with new data that produces classification outputs quickly even though the data has never been recorded by this model before. With these results, the development of detection of glomerular disease can help the performance of the medical team to make decisions more quickly and accurately in dealing with the diagnosis of glomerular disease. In addition, these results can also help the medical team work steps to diagnose chronic kidney disease further.
 S.George et al., “Evidence for SARS-CoV-2 Spike Protein in the Urine of COVID-19 patients,” 2021.
 N.Altini et al., “Semantic segmentation framework for glomeruli detection and classification in kidney histological sections,” Electron., vol. 9, no. 3, 2020, doi: 10.3390/electronics9030503.
 Y. Bengio, A.Courville, andP.Vincent, “Representation Learning: A Review and New Perspective,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 8, pp. 1799–1828, 2013, doi: 10.1109/TPAMI.2013.50.
 J. A.Martín-Navarro, M. J. Gutiérrez-Sánchez, V. Petkov-Stoyanov, P. Justo-Ávila, and R. Ionela-Stanescu, “Focal segmental glomerulonephritis in patients with pulmonary sarcoidosis.,” Nefrol. (English Ed., vol. 33, no. 3, pp. 431–433, 2013, doi: 10.3265/Nefrologia.pre2012.Sep.11753.
 K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., vol. 2016-Decem, pp. 770–778, 2016, doi: 10.1109/CVPR.2016.90.