Autonomous detection of molecular configurations in microscopic images based on deep convolutional neural network

TL;DR

This paper presents a deep learning-based method using Faster R-CNN for high-throughput, autonomous detection and orientation analysis of molecules in microscopic images, achieving high accuracy.

Contribution

It introduces a novel application of Faster R-CNN for multi-scale molecule detection and orientation measurement in microscopic images, enabling high-throughput analysis.

Findings

High detection accuracy with F1 score close to 1

Effective multi-scale detection across 10-50 nm images

Revealed preferred adsorption configurations of molecules

Abstract

In an effort to explore high-throughput processing of microscopic image data, a method based on deep convolutional neural network is proposed. The state-of-the-art computer vision algorithm, Faster R-CNN, was trained for the detection of iron (II) phthalocyanines on Se-terminated Au(111) platform resolved by scanning probe microscopy. The construction of the feature pyramid enables the multi-scale molecule detection in images of different scales from 10 nm to 50 nm. After the detection, the orientation of each molecule is measured by a following program. Based on the statistical distribution of the orientation angles, the preferred adsorption configurations of iron (II) phthalocyanine on the platform are revealed. This method yields high accuracy and recall with F1 score close to 1 after optimization of hyperparameters and training. It is expected to be a feasible solution in the scenarios where autonomous and high-throughput processing of microscopic image data is needed.