Icon for: Geoffrey Harlow


University of California at Riverside
Years in Grad School: 2

Judges’ Queries and Presenter’s Replies

  • May 20, 2013 | 02:57 p.m.

    Nice and interesting results! I have a few questions regarding the analysis code: is the code done, and if yes what is the amount of time needed to analyze a frame compared to the manual time you give in your video presentation? How are you code and computational approach different from other code(s) that may exist?


  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 10:05 p.m.

    Thank you for your interest and kind words. The code is indeed finished, and was used to quantify the results of an assay. A reviewed paper on this work has been published in the IEEE International Symposium on Biomedical Imaging 1 and it was presented in April 2013. If you have MATLAB and the image processing toolbox, a demo is publicly available 2.

    On the average, with our algorithm, the amount of time needed to analyze a single frame is 0.364s. If there are 20 cells in a frame, and a video has 200 frames, this takes 22.2 minutes for our algorithm to analyze a single video. With manual analysis, the duration of the time required to analyze the same video is 14 days (336 hours). We have reduced the amount of time required by a factor of 908.

    Our goal is to create software that can be used by others that does not require a cutting edge computer. The computer used, in the above statistics, is a model from 2008. It is a Dell Optiplex 780, with the following specifications: Microsoft Windows 7 Enterprise, an Intel Core 2 Duo E8500, an Intel Q45 Chipset, 4GB DDR3 1066Mhz, a 250GB SATA 7200 RPM HDD.

    With respect to the algorithms, a conventional approach employs an edge detector (a Gabor filter is one such filter 3). However, there are significant technical challenges in analyzing the videos such as the background texture and noise that would cause false alarms in the detection of microtubules. This prompted us to develop a novel procedure to estimate the background texture and suppress it when edges are detected. To the best of our knowledge, we are the first to automatically quantify microtubule ordering in pavement cells. This procedure is of broad interest to edge detection in the field of image processing and, recently, it has been successfully applied to facial emotion recognition 4.

    1 G. Harlow, A. C. Cruz, S. Li, N. S. Thakoor, A. C. Bianchi, J. Chen, B. Bhanu and Z. Yang, “Automated spatial analysis of ARK2: Putative link between microtubules and cell polarity,” in IEEE International Symposium on Biomedical Imaging, San Francisco, CA, April 2013.

    2 A. C. Cruz. (2013, May 22). Running Time Demo. Available:

    3 Movellan, J. (2002). Tutorial on Gabor filters. Available:

    4 A. C. Cruz, B. Bhanu, N. S. Thakoor, “Facial emotion recognition with anisotropic inhibited Gabor energy histograms,” in IEEE Int’l. Conf. on Image Processing, Melbourne, Australia, September 2013.

  • May 21, 2013 | 09:32 a.m.

    Hi Geoffrey,
    this is a very interesting method. When you say that the mutation resulted in “significant” changes, how big are these changes?

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 11:01 p.m.

    Thank you for your question. The mutation in ark2-1 resulted in changes of morphology and microtubule ordering. It was hypothesized that the mutant ark2-1 caused less interdigitation (fewer lobes and indentations) and caused the microtubules to have a similar orientation (parallel ordering). We computed the p-value of the two populations (wild type vs. ark2-1) to determine if it was statistically significant. It was found that the solidity, Hu’s moment 1 and excess kurtosis of the microtubules were different between the two populations (p<.05).

    For each cell in each frame, the following features changed significantly between the two populations:
    1. Solidity. The ratio of the area of the cell to the convex hull containing the cell. If the cell had no lobes it would fit well into the convex hull. If the cell had many, large lobes, it would not. It was found that p < 0.0016.
    2. Hu moments. We employ Hu’s moments 1, 2 and 8 1. These features quantify the properties of a shape into a scalar for comparison. If two different cells have a similar Hu’s moment, they are of similar shape. It was found that p <0.0006.
    3. Excess kurtosis of the distribution of microtubule orientation. For microtubules, we automatically computed the distribution of the orientation of the microtubules in each cell, and computed the excess kurtosis, a metric quantifying the distance to a normal distribution. If the orientation of microtubules in a cell had a specific preference (pointed in the same direction), they would have an excess kurtosis close to zero. It was found that p < 0.005.

    The above features were all automatically calculated. Please see our paper for further insight 2.

    1 J. Flusser and T. Suk, “Rotation moment invariants for recognition of symmetric objects,” IEEE Trans. on Image Processing, vol. 15, no. 12, pp. 3784-3790, 2006.

    2 G. Harlow, A. C. Cruz, S. Li, N. S. Thakoor, A. C. Bianchi, J. Chen, B. Bhanu and Z. Yang, “Automated spatial analysis of ARK2: Putative link between microtubules and cell polarity,” in IEEE International Symposium on Biomedical Imaging, San Francisco, CA, April 2013.

  • May 21, 2013 | 10:43 a.m.

    Good study! How does Ark2-1 increase microtubule ordering

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 10:56 p.m.

    This is a very important question. It is not known at this time how the ark2-1 mutation mechanistically increases microtubule ordering. Here in Dr. Zhenbiao Yang’s lab at UC Riverside, several experiments are being conducted to reveal the relationship between ARK2-1, microtubule ordering, and several putative intermediary interactors, of which the preliminary results are promising. Of related interest, it has been hypothesized that a homologous gene in root hairs, ARK1, influences root hair tip growth through the reduction of endoplasmic microtubule accumulation 1. However, this specific implication in the study of root hairs is unclear and these two ARKs may in fact function quite differently. We are still investigating this and thank you for this excellent question.

    1 T. Sakai, H. V. D. Honing, M. Nishioka, Y. Uehara, M. Takahash1, N. Fujisawa, K. Saji, M. Seki, K. Shinozaki, M. A. Jones, N. Smirnoff, K. Okada, and G. O. Wasteneys, “Armadillo repeat-containing kinesins and a NIMA-related kinase are required for epidermal-cell morphogenesis in Arabidopsis,” J. Plant, vol. 53, no. 1, pp. 157-171, 2007.

  • May 21, 2013 | 03:35 p.m.

    Hi Geoffrey,
    Very interesting work iwth the ARK2 interactor. Based on the data shown in Figure 3, it appears as if quantity, not orientation of microtubules is affected by the mutation. Can you address this?

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 11:08 p.m.

    Thank you for your question and interest in our work. While there is a potential that there are a larger quantity of microtubules in either population, our Figure 3 actually compares the results between (top) a state-of-the-art approach using Gabor filters to detect the edges and (bottom) the proposed non-classical receptive field procedure. However, it does not compare the difference in quantity of microtubules between two different populations.

    In Figure 3, red indicates TubA (microtubules), whereas cyan indicates where the algorithm detected a microtubule. Figure 3 demonstrates that if you used a state-of-the-art algorithm (top) it would falsely detect many more microtubules. If you zoom in closely you can see that the proposed approach (bottom) detects microtubules more accurately. Their performance was verified against a ground truth segmented by experts in the field. It was found that the proposed approach reduces false alarms by a factor of 2.14.

  • Icon for: Zhaomin Yang

    Zhaomin Yang

    May 21, 2013 | 07:52 p.m.

    This Z. Yang turned out not to be me! Anyway, nice interdisciplinary work. How will the computational scientists contribute besides debugging and technical support in the future?

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 11:29 p.m.

    Another Z. Yang! Thank you for your question and kind comments. As an example of future contributions, our lead computational scientist on this project, Albert Cruz, understands fluorescent gene labeling and how to operate a confocal microscope (and can be called upon to do more than debugging and technical support in the future). Conversely, at the end of my training I will possess the skills necessary to create computer vision programs for segmentation and edge detection.
    This provides a good opportunity to describe my experiences in the Video Bioinformatics IGERT. Ours is an interdisciplinary program that brings together life sciences, computer science, electrical engineering, and bioengineering to tackle video bioinformatics problems. In some cases, collaboration takes place where each party concerned contributes only within the confines of their discipline to the project. Traditionally a biologist might collect the data, then a programmer creates software for the biologist. Both parties miss out on a fuller understanding of their work.
    In contrast, our IGERT has formed a community of research, with each party learning much from the other. While I cannot speak to what occurs in other IGERTs, as a part of our IGERT program we take part in rigorous coursework that educates us in other fields. We are educated in: (1) modeling multi-scale systems in bio-engineering and grant writing, (2) bio-imaging and bio-instrumentation (including the use of confocal microscopy and fluorescent protein labels), (3) medical imaging (including the use of MRI, CT and PET modalities), (4) imaging of live cells such as stem cells, and (5) computer vision and image processing. At the completion of our training, we are competent in these disciplines.
    The work presented here is developed in a team environment where computer scientists learn from biologists and vice versa. This interaction is necessary to develop a deep understanding and appreciation of the biological problems, developing robust computer algorithms, evaluating their effectiveness and developing still more effective and efficient algorithms and associated software to be used by both biologists and computer scientists. Without an effective team it would not be possible to perform this interdisciplinary research. We believe that we have a strong collaborative interdisciplinary team for this project.

  • Further posting is closed as the competition has ended.

Presentation Discussion
  • May 22, 2013 | 02:21 p.m.

    hi geoffrey! interesting work, and the video is clear and well structured! good job, and good luck!

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 10:23 p.m.

    Thank you. As a testament to their support, the fellows, advisors, and PIs in our Video Bioinformatics IGERT provided many helpful recommendations.

  • Small default profile

    Cecilia Donato

    May 22, 2013 | 03:16 p.m.

    Hi Mr. Geoffrey Harlow,
    I think it’s really interesting to see this video, very clear and easy to understand for everybody.
    Cecila M Donato
    Paris, France

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 22, 2013 | 10:15 p.m.

    Wow, all the way from France! Thank you for your kind comments and interest.

  • Icon for: Philomena Chu

    Philomena Chu

    May 23, 2013 | 12:43 a.m.

    Great video! Clear explanation and cool microscopy images showing the differences between mutant and WT phenotypes. I also liked that you collaborated with an electrical engineer to help you with your research.

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 23, 2013 | 01:51 a.m.

    Thanks Philomena, I really liked your project. You also have such a cool name.

  • Icon for: Daniel Jacobs

    Daniel Jacobs

    May 23, 2013 | 12:55 p.m.

    Hey Geoffrey. You did an excellent job on this video. This is excellent work on analyzing the leaf structure. What the significance is on the level of leaf ordering? I see that with more ordering in the mutated samples, there are smaller more circular lobes developed, but how does this effect the plant? I didn’t catch in the video or poster if and how this is better or worse for plant growth. Admittedly, I know nothing about this field so I apologize if this is an obvious question.

  • Icon for: Geoffrey Harlow

    Geoffrey Harlow

    May 23, 2013 | 03:25 p.m.

    Intuitively it seems that well-ordered microtubules produce leaves mechanically weaker to withstanding forces perpendicular to the ordering. We have a strategy to test these hypotheses and expect to relate specific unique mutations we have identified to improvements or declines in mechanical strength.

    As it pertains to circularity, plant leaves displaying more circular cells are easier to pull apart compared to wild-type. One hypothesis is that increased lobe # and lobe length (features our software automatically quantifies) contribute to greater tensile strength, much like folded hands with interlacing fingers are harder to separate. It is also expected that a specific lobe height & width confer optimal mechanical strength to the total leaf to withstand the pressures of the environment.

    It’s a fair question, and I’m not sure the answer is so obvious. I’m very interested in understanding how a single gene mutation can effect cell shape, conferring an overarching effect to the total leaf and ultimately plant survival under harsh conditions.

    We cannot know the full impact of our basic science cell polarity research until the opportunity presents itself to utilize the knowledge in applied science, but one clear inroad is the implications this has on cell strength.

  • Further posting is closed as the competition has ended.