Performing Video Frame Prediction of Microbial Growth with a Recurrent Neural Network

TL;DR

This work demonstrates that PredRNN can forecast future frames of microbial growth in a microfluidics-fluorescence setup involving two Pseudomonas aeruginosa mutants, using 20-frame videos where the first 10 frames predict frames 11–20. By augmenting 48 seven-dynamic-frame videos to 392 twenty-frame samples and training for 50,000 iterations, the authors achieve good qualitative and quantitative agreement in test wells, as shown by MSE and LPIPS alongside population curves and colony-size analyses. The study highlights the potential for autonomous microbiology experiments by enabling edge-device inference to decide whether to continue imaging a given well, and it underscores the need for larger, balanced growth datasets to improve robustness. Overall, the results provide a foundational step toward data-driven, autonomous experimentation in microbiology.

Abstract

A Recurrent Neural Network (RNN) was used to perform video frame prediction of microbial growth for a population of two mutants of Pseudomonas aeruginosa. The RNN was trained on videos of 20 frames that were acquired using fluorescence microscopy and microfluidics. The network predicted the last 10 frames of each video, and the accuracy's of the predictions was assessed by comparing raw images, population curves, and the number and size of individual colonies. Overall, we found the predictions to be accurate using this approach. The implications this result has on designing autonomous experiments in microbiology, and the steps that can be taken to make the predictions even more accurate, are discussed.

Figures (6)

• Figure 1: Brightened snapshots at time steps 1, 4, 7, 10, 13, for one of the videos obtained in WOS:000456223900011 for the 30$\mu$m well.
• Figure 2: Convergence metrics of PredRNNDBLP:journals/corr/abs-2103-09504 training. (a) MSE loss for training images. (b) MSE loss for validation images. (c) LPIPS loss for validation images.
• Figure 3: Qualitative comparison between the groundtruth and predicted frames for four wells in the test dataset. The upper panel represents the groundtruth frames, and the lower the predicted frames. The first image at the left of each panel represents the 11$^{th}$ frame, followed by 13$^{th}$, 15$^{th}$, 17$^{th}$ and 19$^{th}$. The wells are numbered according to their position in the test dataset.
• Figure 4: Quantitative comparison between the groundtruth and the predicted frames for the test wells in Fig.\ref{['fig:fig2']}. (a) average Mean-Squared Error (MSE) for all test wells; (b) average Learned Perceptual Image Patch Similarity (LPIPS) for all test wells. Each dot represents a frame.
• Figure 5: Population curve comparisons from groundtruth and PredRNNDBLP:journals/corr/abs-2103-09504 generated predictions