Small hand-designed convolutional neural networks outperform transfer learning in automated cell shape detection in confluent tissues

Abstract

Mechanical cues such as stresses and strains are now recognized as essential regulators in many biological processes like cell division, gene expression or morphogenesis. Studying the interplay between these mechanical cues and biological responses requires experimental tools to measure these cues. In the context of large scale tissues, this can be achieved by segmenting individual cells to extract their shapes and deformations which in turn inform on their mechanical environment. Historically, this has been done by segmentation methods which are well known to be time consuming and error prone. In this context however, one doesn’t necessarily require a cell-level description and a coarse-grained approach can be more efficient while using tools different from segmentation. The advent of machine learning and deep neural networks has revolutionized the field of image analysis in recent years, including in biomedical research. With the democratization of these techniques, more and more researchers are trying to apply them to their own biological systems. In this paper, we tackle a problem of cell shape measurement thanks to a large annotated dataset. We develop simple Convolutional Neural Networks (CNNs) which we thoroughly optimize in terms of architecture and complexity to question construction rules usually applied. We find that increasing the complexity of the networks rapidly no longer yields improvements in performance and that the number of kernels in each convolutional layer is the most important parameter to achieve good results. In addition, we compare our step-by-step approach with transfer learning and find that our simple, optimized CNNs give better predictions, are faster in training and analysis and don’t require more technical knowledge to be implemented. Overall, we offer a roadmap to develop optimized models and argue that we should limit the complexity of such models. We conclude by illustrating this strategy on a similar problem and dataset.

Accelerating computer vision-based human identification through the integration of deep learning-based age estimation from 2 to 89 years

“…Furthermore, these studies utilize transfer learning, advantageous for achieving rapid and satisfactory results with limited datasets. Ad-hoc CNNs may outperform transfer learning when trained specifically for a particular solution 33 , 34 . Transfer learning, often relying on pretraining with non-medical images, carries the risk of inheriting undesired patterns or biases, potentially diminishing accuracy in addressing specific issues.…”

Inferring the location and orientation of cell divisions on time-lapse image sequences

“…The first stage comprises 128 filters; this number is doubled for each subsequent block. After the 4th block, we use 2 regression heads on the flattened features: one scalar l for the length, and one vector (cos(2 α ), sin(2 α )) for the orientation, which we preferred to the scalar α , as to avoid issues related to 0 and π representing the same orientation, while appearing distant in the classical MSE [10]. We used data augmentation and training protocols similar to the one described for the U-Net.…”