Artifact detection in fluorescence microscopy using convolutional autoencoder

To ensure analytical accuracy in fluorescence microscopy image analysis, robust artifact detection is essential. For large datasets or time-sensitive analyses, automation is advisable, as it not only reduces time and costs but also eliminates human bias and enhances reproducibility. Although artificial intelligence is commonly employed for artifact detection, it is typically limited to recognizing artifact types that have been previously learned, often necessitating large training datasets. This study proposes an approach for an automated detection of previously unseen artifacts without the need for a training set of artifact-laden images. Multiple datasets were assembled using images generated by our surface-based intensity distribution analysis (sFIDA) technology during different experiments. A convolutional autoencoder was trained on a dataset of artifact-free images to reproduce preprocessed images accurately. Artifact-laden images are subsequently detected by computing the difference between the input and output of the model, with increased discrepancies indicating the presence of artifacts. The proposed model is capable of classifying artifacts across different datasets with an average accuracy of 95.5%. Additionally, the model was able to detect unseen artifacts of various types, including differences in cause, structure, size and intensity. The findings demonstrate that convolutional autoencoders provide a lightweight, but effective method for detecting artifact-laden images. While the method was tested only on sFIDA images, its design, which does not rely on an artifact specific training set, suggests potential for use across various microscopy techniques.