Quantifying lymphocyte vacuolization in peripheral blood smears (PBSs) serves as a measure for disease severity in CLN3 disease-a lysosomal storage disorder of childhood-onset. However, thus far quantification methods are based on labor-intensive manual assessment of PBSs. As machine learning techniques like convolutional neural networks (CNNs) have been deployed quite successfully in detecting pathological features in PBSs, we explored whether these techniques could be utilized to automate quantification of lymphocyte vacuolization. Here, we present and validate a deep learning pipeline that automates quantification of lymphocyte vacuolization. By using two CNNs in succession, trained for cytoplasm-segmentation and vacuolization-detection, respectively, we obtained an excellent correlation with manual quantification of lymphocyte vacuolization (r = 0.98, n = 40). These results show that CNNs can be utilized to automate the otherwise cumbersome task of manually quantifying lymphocyte vacuolization, thereby aiding prompt clinical decisions in relation to CLN3 disease, and potentially beyond.
Keywords: CLN3 disease; lymphocyte vacuolization; machine learning; neuronal ceroid lipofuscinosis.
© 2021 The Authors. JIMD Reports published by John Wiley & Sons Ltd on behalf of SSIEM.