A U-Net-based approach for automated segmentation of scanning artifacts in multi-modal cellular AFM images

Abstract

Here we present a method for the automatic segmentation of imaging artifacts in atomic force microscopy (AFM) maps collected on cell surfaces. The method is based on a U-Net convolutional neural network architecture and uses multi-channel AFM data, including height, adhesion, and restored adhesion maps acquired in PeakForce QNM, HybriD, and Ringing modes. The proposed method demonstrated high accuracy in segmenting artifact locations, as confirmed by receiver operating characteristic-area under the curve and PR-AUC metrics. The intersection over artifact, metric introduced in this work, reached a maximum value of 0.95, indicating a nearly complete delineation of areas containing artifacts. The developed approach is an efficient and scalable tool for automating the preprocessing and screening of biological AFM images, which significantly reduces the need for labor-intensive manual processing and minimizes the influence of human bias.