203. Evaluation of CNN methods for computer-assisted morphometry in kidney biopsies from AAV patients

Background: Although machine learning methods have just come out of their infancy, with the first AI-based decision support system for clinical diagnostics of prostate cancer having been approved by US FDA in September 2021, and the focus is on the field of oncopathology, AI techniques have also shown promising results in nephropathology, where AI based methods can exploit the large amount of information given by WSIs (Whole Slide Images). Visual assessments by pathologists don’t make use of all the information due to the huge dimensions of images and consequent time constraints and human perception of complexity is limited, whereas machine learning techniques thrive with large amount of data and are able to find interdependencies that remain unrecognized in visual observation. Moreover, machine learning techniques already achieved performance levels comparable to human capabilities in several visual tasks. Glomerulosclerosis constitutes a critical prognostic marker for CKD progression as is IFTA, but its identification is based on the count of individual discrete structures, glomeruli, rather than an overall visual assessment. As a consequence, it suffers less variability in visual assessment, although still representing a tedious and time-consuming task for pathologists, which could be successfully automated via computer assisted techniques.

Methods: CNN (Convolutional Neural Network) architectures have been implemented and trained on a total of 3120 tiles (512x512 pixels) and finally evaluated on 878 tiles. U-Net architecture has been elected as being successful in several segmentation tasks within the biomedical field due to its remarkable capability to capture features at different resolutions via its contracting and expanding architecture. Consequently, several U-Net derived methods have been implemented and evaluated to assess whether alternative configurations could further improve diagnostic and prognostic fidelity in glomeruli segmentation. Data augmentation techniques have been applied to enhance variability within the training set and models have been evaluated with and without data augmentation.

Results: U-Net based methods show high stability in training as compared to other CNN techniques (e.g. Deep Convolutional Generative Adversarial Networks). Although yielding promising results in terms of performance, metrics results are highly variable across the models due to the large number of possible parameters and configurations, and models often suffer from artifacts within the image prediction. Also, staining techniques seem to affect model performance, thus suggesting that specific stains are to be matched to specific segmentation tasks, which is also reflected by the use of different stains during pathologists’ visual assessment of biopsy specimens.

Similarly, color normalization also needs to be taken into account in order to reduce colors variability across images in the training set, which could be a confounding factor.

Conclusions: Deep learning techniques can be applied for glomeruli segmentation reliably and prove to be able to capture informative features across the image at different resolution levels due to the several layers used in the contracting and expanding path. U-Net derived methods also show promise in terms of tailoring the use of conventional CNN networks towards specific needs, although suffering from the lack of optimization, which well-established methods enjoy. Further research is required to both optimize these methods and assess whether and under what terms medicine could benefit more from them, and to evaluate their use on a larger cohort and on a different training set, to evaluate the potential impact of variability across datasets on performance.

Disclosures: none