2019
Christos Madamopoulos Nishant Mishra Aaron Serianni
Aaron Serianni:
For the past two summers, I have been working with another LLP student Bo Deng under the guidance of Dr. Jason Puchalla through the Princeton Laboratory Learning Program for high school students. In the Puchalla Lab, I am conducting research on the applications of deep convolutional neural networks (CNNs) for biological classification, specifically for juvenile zebrafish identification. Zebrafish have become an important model organism for use in biomedical research, replacing the more traditional use of mice and rats. Their small size and similar features make it hard for humans to differentiate between individual zebrafish. Traditional methods for identifying zebrafish individuals include tagging, clipping fins, injecting dye subdermally, and implanting RFID tags. All these methods can affect the results of biomedical studies due to their invasiveness. Hence, the most common method is to take a picture of a zebrafish and compare by eye it to a chart of previously taken photos of each zebrafish, which is time consuming and inaccurate, especially when using a large sample size. Using CNNs, we created an alternate zebrafish recognition pathway which is non-invasive and less labor intensive than by human eye. Convolutional neural networks are a powerful and versatile form of machine learning, based on the visual cortex in biological brains. With the recent rise of powerful graphic processing units (GPUs) in the past decade, CNNs have become very popular in the field of image recognition and classification, resulting in the best performances on numerous datasets. We chose to use CNNs for these reasons and applied them for the recognition of animal individuals. However, a major drawback of neural networks is their “black box” nature, which we probed and overcame through various analysis techniques. We created a tank-insertable photo studio, allowing for a non-invasive way of isolating an individual zebrafish in a small area against a white background. Using the studio, we created a dataset of over 4000 high-quality images for 5 different juvenile zebrafish, taken over the timespan of 19 days for robustness. For the project, a machine learning pathway was created using the Tensorflow and Keras APIs, written in Python. This includes image preprocessing, data augmentation techniques, and the neural network itself. We opted to use a modified version of Google’s InceptionV3 CNN, using pertained weights created from the ImageNet dataset. The CNN model was then trained on the Princeton Physics Department’s GPU computer cluster. Over multiple runs, the model was able to achieve 99-100% accuracy on testing sets, showing the pathway as a viable method for zebrafish recognition. In addition, we investigated the robustness of the model over multiple weeks, and while the accuracy of the model decreased to a minimum of 93%, the training data can be supplemented with additional up-to-date images. We investigated the relevance of various features of the zebrafish images to characterize the success of the CNN through. To do so, I created a class activation heatmap visualization, and used targeted image modification of the training set. Through these techniques, we found that the body of zebrafish had the most significant impact for recognition, with the fins and tail affecting the model less, and the head being negligible. I also did a comparison of the CNN model on our dataset to other machine learning classifiers, built through the scikit-learn library. In addition, I also tested the smaller UNet CNN model, which has been used in other animal recognition projects. In all cases, the InceptionV3 model was able to significantly outperform the other classifiers. From our research, we were able to develop and demonstrate a robust identification protocol for free-swimming juvenile zebrafish, and explore the effectiveness of this technique and the nature of the deep CNN model. Currently, we are finishing a paper on this project, which will soon be submitted for publication.Nishant Mishra:
A protein's capability to perform specific biological functions depends on whether or not it can properly aggregate/disaggregate with other proteins. This aggregation, however, can be incredibly complex and has become a widely studied field that has enabled researchers to obtain a stronger understanding of disorders such as Alzheimer's and Parkinson's. One way in which this aggregation can be considered is through the observation of the macromolecular interactions between individual particles, otherwise known as single-particle kinetics. To take the measurements associated with this field of study, it is useful to have access to an environment where the rate and direction of fluid flow can be manipulated to simulate various macromolecular interactions.Over the summer, I worked on the create such an environment by developing a microfluidic flow system. To facilitate the fluid flow, I networked three NE- 500X2 syringe pumps (New Era Pump Systems), and connected them to a single computer port. To qualitatively log the stability of the fluid flow, I filled each of the syringes in the pumps with fluorescent bead solution, lead them into channels in PDMS chips, and observed the flow of beads in the channels under a Nikon Ti-S Microscope.
To send commands to the three syringe pumps in my system, I coded a user interface using LabVIEW, a system-design platform used for visual programming. Using LabVIEW allows users to modify the code to integrate triggers for other lab hardware, as well as to couple the code with related software, such as Measurement Studio and modeFRONTIER. Along with being able to run the initial configuration procedures, the interface is able to start/stop individual pumps, update the rate while they are running, and set timers to stop each pump after they have infused/withdrawn a certain amount.
Alongside the microfluidic system development, I also worked on a set of MATLAB functions that simulate the distributions of the fluorescent bursts seen in a camera image of various excitation fields (exponential, power, etc.).
Currently, I am working on a machine learning project centered around the customization of deep learning algorithms to sort depth map images. The results will be used to analyze whether the added data from depth maps can improve substantially improve the effectiveness of identification done by neural networks.
Christos Madamopoulos:
In this project, the response of ferrofluid, and in particular its shape, under different boundary conditions (e.g., container shape and/or flexibility) and under the influence of a uniform magnetic field are studied. The understanding of the ferrofluid response under these boundary conditions can aid in the understanding of how boundary conditions affect an biological development, e.g. how the boundary conditions can change the shape of the embryo’s body/body parts. In these experiments, the boundary conditions are controlled using:I. A pair of electromagnets in a true Helmholtz configuration. This pair produces a uniform magnetic field. The field strength is controlled by the electrical current flowing through the coils of the electromagnets. Several mechanical assemblies and configurations were designed and tested to develop the optimum experimental setup. The goal was to achieve a strong magnetic field with extended uniform distribution. A constant current power supply was used to accurately control the electric current. The field strength was controlled by controlling the electrical current that flowed through the coils with a constant current power supply. In order to achieve an extended and uniform magnetic field, large electromagnets (e.g., 15.24 cm in diameter) were built, with appropriate gage wire to withstand high current. The size of the electromagnets and the appropriate current source have been configured to provide the required magnetic field. II. Different shape and materials were used to loosely contain the ferrofluid. This is a very challenging task, since the goal is to emulate an infinite pool of ferrofluid while using a finite amount of that fluid. The materials used to contain the ferrofluid must be non wettable by the ferrofluid (i.e., a contact surface angle of at least 90°). This would guarantee that its surface tension with the wall of the container does not change the ferrofluid pattern formations. The material must also be flexible enough to change its shape under the force of the ferrofluid, but also rigid enough to form and sustain corners. Currently, we are exploring different material options and configurations to develop such a container.Future works relates with the application of different strength magnetic field in the infinite pool of ferrofluid and the quantification of the effect, e.g., the critical magnetic field strength where the ferrofluid starts forming into specific patterns. Under this context, we will investigate techniques to create (a) a small pool of ferrofluid in a larger extent magnetic field and (b) a large pool of ferrofluid in a smaller extend magnetic field, with a goal to emulate a close as possible the infinite pool condition.