An image segmentation method based on the spatial correlation coefficient of Local Moran’s I

  1. Laboratory of Thalamus Research, Institute of Experimental Medicine, Budapest, 1083 Szigony u 43
  2. Department of Anatomy, Histology and Embryology, Semmelweis University, Budapest, H-1094 Budapest, Tűzoltó u 54
  3. Laboratory of Cellular Neurophysiology, Institute of Experimental Medicine, Budapest, 1083 Szigony u 43
  4. Novarino Group, Institute of Science and Technology, Austria, A-3400 Klosterneuburg
  5. Sensory Circuits and Neurotechnology Laboratory, Francis Crick Institute, 1 Midland Road, NW1 1AT, London, UK

Peer review process

Not revised: This Reviewed Preprint includes the authors’ original preprint (without revision), an eLife assessment, public reviews, and a provisional response from the authors.

Read more about eLife’s peer review process.

Editors

  • Reviewing Editor
    Eunji Cheong
    Yonsei University, Seoul, Korea, the Republic of
  • Senior Editor
    John Huguenard
    Stanford University School of Medicine, Stanford, United States of America

Reviewer #1 (Public Review):

The study describes a new computational method for unsupervised (i.e., non-artificial intelligence) segmentation of objects in grayscale images that contain substantial noise, to differentiate object, no object, and noise. Such a problem is essential in biology because they are commonly confronted in the analysis of microscope images of biological samples and recently have been resolved by artificial intelligence, especially by deep neural networks. However, training artificial intelligence for specific sample images is a difficult task and not every biological laboratory can handle it. Therefore, the proposed method is particularly appealing to laboratories with little computational background. The method was shown to achieve better performance than a threshold-based method for artificial and natural test images. To demonstrate the usability, the authors applied the method to high-power confocal images of the thalamus for the identification and quantification of immunostained potassium ion channel clusters formed in the proximity of large axons in the thalamic neuropil and verified the results in comparison to electron micrographs.

Strengths:
The authors claim that the proposed method has higher pixel-wise accuracy than the threshold-based method when applied to gray-scale images with substantial noises.

Since the method does not use artificial intelligence, training and testing are not necessary, which would be appealing to biologists who are not familiar with machine learning technology.

The method does not require extensive tuning of adjustable parameters (trying different values of "Moran's order") given that the size of the object in question can be estimated in advance.

Weaknesses:
It is understood that the strength of the method is that it does not depend on artificial intelligence and therefore the authors wanted to compare the performance with another non-AI method (i.e. the threshold-based method; TBM). However, the TBM used in this work seems too naive to be fairly compared to the expensive computation of "Moran's I" used for the proposed method. To provide convincing evidence that the proposed method advances object segmentation technology and can be used practically in various fields, it should be compared to other advanced methods, including AI-based ones, as well.

This method was claimed to be better than the TBM when the noise level was high. Related to the above, TBMs can be used in association with various denoising methods as a preprocess. It is questionable whether the claim is still valid when compared to the methods with adequate complexity used together with denoising. Consider for example, Weigert et al. (2018) https://doi.org/10.1038/s41592-018-0216-7; or Lehtinen et al (2018) https://doi.org/10.48550/arXiv.1803.04189.

The computational complexity of the method, determined by the convolution matrix size (Moran's order), linearly increases as the object size increases (Fig. S2b). Given that the convolution must be run separately for each pixel, the computation seems quite demanding for scale-up, e.g. when the method is applied for 3D image volumes. It will be helpful if the requirement for computer resources and time is provided.

Reviewer #2 (Public Review):

Summary:
The manuscript by David et al. describes a novel image segmentation method, implementing Local Moran's method, which determines whether the value of a datapoint or a pixel is randomly distributed among all values, in differentiating pixel clusters from the background noise. The study includes several proof-of-concept analyses to validate the power of the new approach, revealing that implementation of Local Moran's method in image segmentation is superior to threshold-based segmentation methods commonly used in analyzing confocal images in neuroanatomical studies.

Strengths:
Several proof-of-concept experiments are performed to confirm the sensitivity and validity of the proposed method. Using composed images with varying levels of background noise and analyzing them in parallel with the Local Moran's or a Threshold-Based Method (TBM), the study is able to compare these approaches directly and reveal their relative power in isolating clustered pixels.

Similarly, dual immuno-electron microscopy was used to test the biological relevance of a colocalization that was revealed by Local Moran's segmentation approach on dual-fluorescent labeled tissue using immuno-markers of the axon terminal and a membrane-protein (Figure 5). The EM revealed that the two markers were present in terminals and their post-synaptic partners, respectively. This is a strong approach to verify the validity of the new approach for determining object-based colocalization in fluorescent microscopy.

The methods section is clear in explaining the rationale and the steps of the new method (however, see the weaknesses section). Figures are appropriate and effective in illustrating the methods and the results of the study. The writing is clear; the references are appropriate and useful.

Weaknesses:
While the steps of the mathematical calculations to implement Local Moran's principles for analyzing high-resolution images are clearly written, the manuscript currently does not provide a computation tool that could facilitate easy implementation of the method by other researchers. Without a user-friendly tool, such as an ImageJ plugin or a code, the use of the method developed by David et al by other investigators may remain limited.

Author Response

We are grateful for the constructive comments of the reviewers and for the succinct assessment of our work by the editors. Here we provide a brief summary of our response to answer the major criticism of our reviewers. We will give a detailed point-to-point response soon when we upload a revision of our paper.

  1. The MATLAB code for the spatial autocorrelation analysis is now freely available at the following site: : https://github.com/dcsabaCD225/Moran_Matlab/blob/main/moran_local.m If any question arises during its implementation, please contact Csaba Dávid (david.csaba@koki.hu)

  2. Concerning the computer resources and times required to perform Moran’s I image analysis, here we provide a brief description of the hardware and the calculations for images with different sizes.

Hardware used for performing the analysis:

Intel(R) Xeon(R) Silver 4112 CPU @ 2.60GHz, 2594 Mhz, 4 kernel CPU, 64GB RAM, NVIDIA GeForce GTX 1080 graphic card.

MATLAB R2021b software was used for implementation.

Computation times are shown in Author response table 1.

Author response table 1.

  1. In response to the comment:

“While the method's avoidance of AI training appeals to those lacking computational know-how and shows improved accuracy over basic threshold-based techniques, there are valid concerns regarding its performance in comparison to advanced methodologies”.

Comparison of Moran’s I image analysis with AI based segmentations raises conceptual problems which will be addressed in detail in the revised version. Briefly, the basis of AI based analyses is that the ground truth is known and using a large teaching set AI learns to extract the relevant information for image segmentation. In several cases, however (like protein distribution in the membrane) the ground truth is not known and cannot be easily determined by any single observer. Defining spatial inhomogeneities in protein distribution, differentiating proteins involved vs not involved in clusters is highly subjective. Indeed, our analysis showed the 23 expert human observers varied hugely in establishing the boundaries of a protein cluster. As a consequence, establishing and using a teaching set would be highly contentious in these cases. In an average laboratory setting generating a teaching set using hundreds of images examined by two dozen people would not be impossible but not really plausible. The beauty of Moran’n I analysis is that it is able to extract the relevant signals from an image generated in different, often noisy condition using a simple algorithm that allows quantitative characterization and identification of changes in many biological and non-biological samples.

  1. Howard Hughes Medical Institute
  2. Wellcome Trust
  3. Max-Planck-Gesellschaft
  4. Knut and Alice Wallenberg Foundation