Locating critical events in AFM force measurements by means of one-dimensional convolutional neural networks

Abstract

Atomic Force Microscopy (AFM) force measurements are a powerful tool for the nano-scale characterization of surface properties. However, the analysis of force measurements requires several processing steps. One is locating different type of events e.g., contact point, adhesions and indentations. At present, there is a lack of algorithms that can automate this process in a reliable way for different types of samples. Moreover, because of their stochastic nature, the acquisition and analysis of a high number of force measurements is typically required. This can result in these experiments becoming an overwhelming task if their analysis is not automated. Here, we propose a Machine Learning approach, the use of one-dimensional convolutional neural networks, to locate specific events within AFM force measurements. Specifically, we focus on locating the contact point, a critical step for the accurate quantification of mechanical properties as well as long-range interactions. We validate this approach on force measurements obtained both on hard and soft surfaces. This approach, which could be easily used to also locate other events e.g., indentations and adhesions, has the potential to significantly facilitate and automate the analysis of AFM force measurements and, therefore, the use of this technique by a wider community.

Application of self-organizing maps to AFM-based viscoelastic characterization of breast cancer cell mechanics

“…Most of the neural network or machine learning approaches that have been applied to AFM force spectroscopy data on biological materials are based on supervised methods. Recently, three different approaches have been provided in the literature: Using machine learning to classify the quality of curves 41 , training a machine learning algorithm with curve shapes 44 , or using curve characteristic properties as inputs for supervised training 43 , 45 , 60 . Such approaches have shown promising results in correctly classifying cancer cells from healthy cells, as well as cancerous tissue.…”

“…Recently, machine learning (ML) and neural network algorithms have been applied to surface probe microscopy data 38 , including image segmentation 39 , automatic data processing 40 – 43 , cancer cell classification and progression grade evaluation 44 – 48 . These approaches were performed in a supervised way, with manual classification of a data set to train the ML algorithms that were applied to classify data.…”

Application of self-organizing maps to AFM-based viscoelastic characterization of breast cancer cell mechanics

“…Most of the neural network or machine learning approaches that have been applied to AFM force spectroscopy data on biological materials are based on supervised methods. Recently, three different approaches have been provided in the literature: Using machine learning to classify the quality of curves 25 , training a machine learning algorithm with curve shapes 28 , or using curve characteristic properties as inputs for supervised training 27,29,46 . Such approaches have shown promising results in correctly classifying cancer cells from healthy cells, as well as cancerous tissue.…”

“…Recently, machine learning and neural network algorithms have been applied to surface probe microscopy data 22 , including image segmentation 23 , automatic data processing [24][25][26][27] , cancer cell classification and progression grade evaluation [28][29][30][31][32] . All these approaches were performed in a supervised way, consisting of a training set with defined classification, that could be used to classify data after training.…”