Teaching Image Processing and Optical Engineering to University Biology Students

Microscopes are an essential tool in biology. Today's microscopes comprise optical components and a digital image sensor, with much of the analysis of captured images processed with computers, for example to identify, count, and measure cell features. Commercial and open source software exists to perform many of these tasks, such as ImageJ (version 1.53t, Wayne Rasband, National Institute of Mental Health, Bethesda, MD). Scientists continually adapt and enhance existing microscopy hardware and image processing tools to expand their compatibility. For instance, light-sheet microscopy was adapted to monitor neurological activity of live zebrafish (1), and fluorescent microscopy was enhanced to exceed the resolution limit set by the diffraction of light (2). These “super-resolution” methods are achieved by teams with a broad range of skills, not only in biology, but in hardware, software, and “soft skills,” such as teamwork and leadership (3–5).

The value of educating students in at least one field with a workable level of understanding in another field is well recognized in engineering, computer science, and management (6–10). In addition to mastering academic knowledge and participating in laboratory courses that teach hands-on technical skills like pipetting, cell culturing, and spectroscopy, students need to develop communication and interpersonal skills to be productive members of a team (11). Students can learn these soft skills with cooperative learning, characterized by students working in small groups with shared goals, collaborative behavior, positive interdependence, and individual and group accountability and responsibility (12). For engineering students, a capstone project is an opportunity to perform independent group research, often with an industrial partner, providing a real-world context for the project and preparing them for the transition from academia to industry (13).

Course-based undergraduate research experiences (CUREs) have emerged as an effective approach to expand research skills for undergraduates with many benefits to students, including enhancing self-confidence (14), increasing diversity (15), and engaging in authentic inquiry-based research (16). Often, student research projects originate from outside the student's home institution (17). We implement a CURE laboratory-based learning course at San Francisco State University where students use microscopy and image processing to address a research question motivated by their scientific curiosity. Student projects often involve collaboration with researchers from Center for Cellular Construction (18) institutions, where collaborators help students formulate projects, share code, and provide datasets developed from their work.

The paper is organized as follows: section II describes the pedagogical principles of our Optical Engineering Laboratory syllabus which is composed of 3 sections that take place in one semester: (section II.A) a self-paced online bootcamp to teach basic programming skills necessary for image processing, (section II.B) an image processing pipeline, and (section II.C) a lab-based hands-on section in which each student builds a holographic microscope and performs original research, often with their microscope. In section III, we review the materials and methods used to teach the 3 parts described in section II. In section IV, we present the results from several student research projects. In section V, we discuss our experiences and observations from teaching the course and students' feedback from pre- and postcourse interviews. We provide a summary of the motivation and intended effect of the course on students' education and professional careers in section VI.

Our approach to the Optical Engineering Laboratory syllabus is threefold: first, teach students the basic programming skills they will need to understand and write image processing code; second, teach students image processing coding with a pipeline and dataset developed for a real research problem; third, engage students in hands-on learning of basic optical theory and engineering skills by building a holographic microscope and using it to conduct original research projects.

We selected Python (version 3.7, Python Software Foundation, Beaverton, OR) as the programming language because it is one of the most common open source languages used in science disciplines, it is easy to learn, and it has many versatile libraries covering many disciplines (19). However, many of our students are not proficient in Python and are new to image processing; therefore, we provide a self-study “bootcamp” to teach the basics of Python programming and image processing. The self-paced instructional modules are run by the students on Google's Colaboratory (Colab) platform (version 2022/6/10, Google, Mountain View, CA), a web-based development environment for Python with instructional material and code exchanged on Google Drive, a cloud-based file storage and sharing platform (20). The modules are publicly available in a GitHub repository (21). Each module is a self-contained Jupyter notebook (release 4.10.0, Jupyter community; https://jupyter.org/) with text and images explaining the purpose of the module, links to papers, videos covering the topic, and a sequence of code segments and their output to teach programming concepts and methods. The Colab platform is well suited for self-paced learning because all the content is available online. No software needs to be installed on the student's laptop, avoiding configuration and device-specific incompatibility problems often experienced when loading an integrated development environment and libraries on a personal computer. The notebooks provide the experience of an interactive textbook, combining text and links to videos and other references, with code, images, and plots that can be run and modified by the student. The modules focus on (a) Python skills commonly used in image processing, including representing and manipulating images with the NumPy library (version 1.21.5) and plotting with the Matplotlib library (version 3.5.1) (22), and (b) image processing pipeline components from OpenCV (version 3.4.2, OpenCV team, Palo Alto, CA) and the scikit-learn library (version 1.0.1) (23).

The image processing pipeline gives students a practical understanding and sufficient skills to use and write code to process the types of images they may encounter in their professional careers. To do this, we use an image processing pipeline we developed to detect and classify plankton (24, 25). The pipeline is composed of components that receive and output CSV (comma-separated value) files, allowing each component to be studied and modified by the student for their research needs. The pipeline is written entirely in Python using the OpenCV library, which students loaded onto their personal laptops with the Anaconda distribution tool (version 2.1.4) (26). The installation is more complex than using Colab but provides several advantages. Code runs much faster on a laptop, speeding up holographic reconstruction, detection, and tracking. Video can stream from the microscope to the laptop, providing real-time feedback, which is essential for adjusting camera exposure settings. Local running code can access the mouse and keyboard, enabling students to write graphical user interfaces (GUIs) to control image processing functions.

Use of the image processing pipeline as a component of the syllabus template provides 3 benefits: (a) a logical progression for the course, because the pipeline performs a sequential series of operations on objects in the image; (b) reuse of code and datasets already developed for research; and (c) teaching students practical skills directly applicable to research and providing collaboration, internship, and employment opportunities.

The pipeline covers a considerable amount of content from image capture to object classification. It was originally developed and taught as a 1-semester lecture course. When used in the Optical Engineering Laboratory course, we teach the Detect and Track components, and leave the remaining pipeline components optional for those students who have sufficient Python programming background and a need for their research project.

Rapid innovations in computing and technology have greatly reduced the cost and increased the accessibility of these tools for biophysical research. To thrive in their scientific careers, the next generation of biophysicists will need to have computational skills, and an understanding of engineering and 3-dimensional (3D) design. Future scientists need to nurture their creativity and problem-solving skills that will drive their future career success. We designed CUREs for students wherein they learn and apply research and engineering skills to answer student-driven research questions. Our pedagogical approach is modeled after curricula designed to promote inclusivity in computing (27) and provides an affordable and scalable model to teach engineering and coding to biology and biochemistry majors.

Laboratory courses serve a multitude of purposes for undergraduates (28). They teach the necessary hands-on skills essential for biologists who want to engage in clinical or wet bench research. They also present a variety of learning styles, including visual, auditory, kinesthetic, tactile, group, and individual learning (29). Building an instrument and using it to collect data provides perceptually rich and realistic educational experiences (30). Additionally, incorporating computer science and engineering experiences into a laboratory course for biology majors is an opportunity to expose more women to engineering, because 3 times as many female graduates have bachelor's degrees in Biology than in Engineering and Computer Science (31). Introducing hands-on engineering experiences to undergraduate women can help decrease the gender bias around tools and machines (32–34).

In the first week of the syllabus (Table 1), students build and test a low-cost holographic microscope (Fig 1) and install Python and the OpenCV library on their personal laptops with Anaconda (35). We start this process the first week because sometimes installation problems are encountered because of the different operating systems (e.g., Windows, Mac, Linux). Code can also use functions and features that are depreciated (no longer supported) with new versions of libraries. We found the question-and-answer website Stack Overflow (36) a good source of assistance in these matters. With a peer-mentored and self-driven learning methodology, students learn the fundamentals of image processing in Python with the Colab notebooks in week 2. In week 3, students apply the skills learned in the first 2 wk to detect and track plankton. They then compare plankton features to see how different species can be distinguished quantitatively. For most of our students, all the material is new and novel.

Nearly all the students have never drilled, soldered, or modified electronics. Most have limited coding experience in Python or R, and for some this is their first coding experience. To support these students, peer mentors and online resources help students better understand the material. Importantly, students learn and develop strategies that work for them to learn how to seek answers to overcome challenges. Additionally, to providing a support framework of peer learning, it also helps promote self-efficacy that drives success in the independent research project that wraps up the remainder of the course.

In weeks 4 and 5, students are taught research methodologies and how to write a scientific paper. Students perform literature searches of potential projects. These discoveries are shared, discussed, and refined. Most projects are driven by student interest in scientific questions of global importance, such as microplastics, chemical pollution, climate change, and artificial intelligence. In this section, students come up with a testable hypothesis based on a novel question, or they focus on a need-based research question with the use of sound scientific methods.

For the remainder of the course (weeks 6–15), students begin their independent research project that was proposed in the previous week. Students build, modify, and design new instrumentation to research their question. They collect initial data and determine how to progress in their research. These projects are usually done in groups, but sometimes students elect to work independently. A fundamental part of this project is that each student takes ownership of their project regardless of whether they are part of a team or work independently. Each week, data are collected and shared with the class and feedback is given. Often, several projects using multiple approaches address a larger research question. Some examples include: (a) How do small amounts of chemical pollutants affect plankton? (b) What is the effect of microplastics on plankton? (c) In what ways can we enhance the research performance of the microscope? (d) How can we classify plankton? In these cases, students can work together on one project or break the project into several complementary subprojects, negotiated during the research planning phase of the course syllabus (weeks 4 and 5; Table 1). For case 1, one student studied the movement of plankton when subjected to common chemicals used in the home and garden (disinfectants and pesticides), another focused on monitoring the natural variation of plankton count and shape sampled from a pond with a net they built, and a third worked on automating sample collection (reported here). For case 2, the students decided to work together to distinguish natural and synthetic microfibers, defining and dividing the tasks among the team (reported here). For case 3, students created several projects to improve the microscope, including measuring resolution under different conditions, exploring 2 approaches to changing magnification, and creating a GUI to improve usability, all reported here. For case 4, 2 students decided to work together, dividing up their work. One student curated plankton sample images, preparing them for feature extraction, and the other performed unsupervised classification, modifying the image processing pipeline code we provided to support student research projects. To demonstrate how a big project can be distributed among many, we had every class member use their holographic microscopes to capture images of plankton (131020, Carolina Biological Supply, Burlington, NC, USA) and provide them to the student curating plankton samples, who learned about the challenges of receiving data (images) of varying quality from multiples sources. In addition to fostering collaboration within and between teams, students were encouraged to consult with faculty outside the classroom and at other institutions. For example, one student worked in a university research lab studying Caenorhabditis elegans and brought in samples to view with their microscope. These activities help students realize that scientific progress is a community effort, and they are part of this community. The course culminates with students giving a final presentation of their research at a departmental seminar.

The following are the 3 syllabus components of the 1-semester Optical Engineering Laboratory course.

All students participate in a bootcamp to learn fundamental skills in image processing. The bootcamp consists of interactive Python scripts to teach Python and OpenCV library skills necessary for image processing written in Jupyter scripts (https://github.com/CCCofficial/Python_Image_Processing_Bootcamp).

• (a) Introduction to Python (all students)

• (b) Working with images (all students)

• (c) Detecting objects in images (all students)

• (d) Tracking detected objects in videos (all students)

• (e) Extracting features from objects (optional, based on research project needs)

• (f) Classifying objects by feature (optional, based on research project needs)

The scripts provide step-by-step demonstrations of applying Python and OpenCV functions. Some include questions for students, with worked out answers in separate scripts.

The image processing pipeline (Fig 2) is available for students whose research projects require tasks like feature extraction and machine learning classification. The image processing pipeline components cover image capture through object classification. Learning and using the Detect and Track component code is mandatory for all students. Students can watch instructional videos and review code on any other components they need for their research project. The videos are from a 1-semester Image Processing Pipeline lecture course we previously taught and recorded with Zoom during the COVID-19 pandemic, edited into 5–10 min segments and organized into 3 YouTube Playlists. For the Optical Engineering Laboratory course, we made the Detect and Track pipeline components mandatory and the remaining components optional according to the needs of the student's research project.

• (a) Image Processing with Python and OpenCV (37–45).

  An introduction to image processing for biologists. Provides an overview of detection, tracking, feature extraction, unsupervised and supervised classification, and the use of a confusion matrix to investigate classification performance.

• (b) Detecting Objects Using Python and OpenCV (46–51). How to use Python and OpenCV functions to detect and track objects in a video. Explains how images are stored in data arrays, how to convert color images into binary images, and how to detect objects by their contour.

• (c) Unsupervised Clustering with Python and Scikit-learn (52–58). How k-means clustering is performed to classify objects automatically, including a review of working code.

The slides used to teach image processing components and holographic reconstruction referenced in Table 1 are publicly available in GitHub repositories (59). The Python code used to teach and implement image the processing of pipeline components, holographic reconstruction, and other techniques to support student research projects is summarized in Table 2. The table includes links to public GitHub repositories that contain the listed code. We shall now discuss how these materials are used in the course syllabus.

Referring to Table 2, the DetectTrack repository contains Python code to detect and track multiple objects in a video, including 2 GUI programs written by a student (co-author Salma Ahmed) to select and save objects from images and videos. A short video of plankton (blep_14sec.mp4), captured with a lensless microscope with a single blue LED located 100 mm from the image sensor, is included for testing. The light source is not coherent enough to produce an interference pattern, so reconstruction is not required, albeit the resolution is lower than the holographic microscope.

The Holo1 repository contains a collection of cropped holographic images of plankton and microfibers (goldHolo.zip) and Python code to perform reconstruction and feature extraction. A paper is included (piClassifer.pdf) that evaluates the feature set performance and run time of 13 classifiers on a Raspberry Pi 3 computer. One student research project (Plankton Classification) combined elements of the Feature.py and Cluster.py programs to perform unsupervised classification. Objects can be at different Z planes, so each must be reconstructed separately. Detecting and tracking holograms of objects, particularly small objects, is often difficult because the fringes spread out the object, reducing the contrast. To address this problem, the program darkestPixelVideo_5.py in the Holo1 repository reconstructs an entire video frame at several Z planes, then selects the darkest pixel from the set of reconstructed Z planes, darkening every object, making them easier to detect.

The Holo2 repository contains Python code and files of holographic video and images and of reconstructed images. Instructions are provided to convert the Raspberry Pi into a USB microscope (USB_CAM.docx), eliminating the need for a video monitor, keyboard, mouse, and power supply. Programs are provided to reconstruct holographic images (findZ.py) and video frames (holoVideoReco.py) and demonstrate how reconstruction works (examineReco.py and phaseAbbScan.py). Several programs are useful utilities and functions (renameFiles.py, keyboard.py, and reco.py). Two programs play notes and chords, created to support a student's research project to sonify plankton by size and location.

Unlike conventional microscopes that require critical alignment of multiple optical components (60), the holographic microscope has only 2 components: an inexpensive laser that is hand aligned above an image sensor (Fig 3). Focusing is performed mathematically by using Fourier transforms (61) applied at the height of the object (Z) over the image sensor (Fig 4). The holographic microscope is based on Dennis Gabor's design published in 1948 (62), wherein he describes how it is possible to eliminate an objective lens and record on photographic film the amplitude and phases of an interference pattern produced by a small object in a field of coherent electromagnetic waves. Gabor used the combination of a mercury arc lamp, narrow-band green filter, and pinhole to create a coherent light source. In our version, we use a red laser (635 nm) and image sensor accessed by removing the lens of a Pi V1 camera (Raspberry Pi Ltd., https://www.raspberrypi.com/documentation/accessories/camera.html), to capture the interference pattern (Fig 4b). However, the red laser is too bright and coherent, saturating the image sensor and filling the image with speckle, a random interference pattern produced by coherent light diffusely reflecting off surfaces. To address this, we use a resistor (330 ohms; Fig 3) to limit the current below the lasing threshold current, so it operates not as a laser, but as a tiny (∼5 μm × ∼100 μm) LED (63). The small size approximates a point light source, emitting continuous light that is acceptable to the image sensor and spatially coherent enough to produce an interference pattern without generating unwanted speckle.

Most of the parts are sourced directly from China (Table 3) and are very inexpensive but can take up to 1 mo to arrive in the United States. The following common accessories are needed to construct the microscope: hot glue gun, soldering iron, diagonal cutters, wire strippers, utility knife, Dremel tool, painter's tape, hand drill, wire, safety googles, and aquarium-safe silicone sealer. A mini-HDMI to HDMI cable is needed to connect the microscope to an external monitor. A USB mouse and keyboard are also needed to control the Pi computer.

The Optical Engineering Laboratory course was offered 4 times between 2019 and 2022: 3 times in-person and 1 time remotely because of Covid-19. For clarity, we will separately describe these 2 modalities.

To optimize class time for construction and research, students reviewed video instructions at home that demonstrate the entire build process (64–67). In class, each student was provided a kit containing the components listed in Table 3 and access to a hand drill (e.g., a Dremel tool), soldering iron, hot glue gun, black silicone sealer, safety glasses, and a few hand tools. The microscope was constructed in a plastic box, oriented upside down, with the thin lid on the bottom to reduce water penetration and simplify construction (Fig 5, right). A sample holder was fabricated by drilling a 3/4-inch (2 cm) hole in 2 polyvinyl chloride (PVC) cards and gluing a 1-inch-diameter (2.5 cm) glass coverslip between the 2 cards with aquarium-safe silicone sealer (Fig 6d), which takes 24 h to dry. This construction created a well to hold a liquid sample. A 3/4-inch (2 cm) hole was drilled in the top of the inverted plastic box, and the sample well was glued above the hole, after determining that the underside of the glass was perfectly clean, determined by operating the microscope. The molded lens of the camera was carefully removed with diagonal cutters to expose the image sensor. A notch was cut in the side of the box with the Dremel tool to provide access to connectors on the side of the Raspberry Pi computer. An SD card containing the free, open source, Linux-based Raspberry Pi OS (version 5.10, Raspberry Pi Foundation, Cambridge, UK) was inserted into the Raspberry Pi, and the Pi board was attached to the plastic case lid with 4 M2.5 (metric) nylon screws. The laser collimating lens was removed and inserted into a 1/8-inch (0.3 cm) hole drilled in the center of a 35-mm square black plastic end cap. The lens was removed to reduce the light intensity that would otherwise saturate the image sensor. Removing the lens also spreads the beam about 18°, simplifying optical alignment with the image sensor. Wires from the laser were passed through a 1/16-inch (1.6 mm) hole in the top of the case and soldered to the 5V and ground pads of the Pi, with a 330-ohm resistor in series to limit current, providing continuous power to the laser. The current was below the lasing threshold (63), operating the laser as a point-light source LED, because lasing would produce too much speckle and brightness. In the final alignment step, the student would turn on the system and slide the cap until a bright uniformly illuminated image was observed on the monitor, then fix the cap location to the case with double-stick foam tape. The assembled microscope is displayed in Figure 6f.

To capture holograms, a sample was placed in the sample well (Fig 6d), the cap was closed, and a 30-s video was recorded by the raspivid utility included in the Raspberry Pi OS (68). Students transferred holographic videos to their laptop with a USB flash drive. The Raspberry Pi–native h.264 video format was converted to MP4 format by the FFmpeg utility (version N-67100-g6dc99fd, FFmpeg team; https://ffmpeg.org) to enable random frame access under program control. We wrote an application for the students to view the video frame-by-frame, select a plankton of interest, then manually step through reconstruction distance values (Z) to select and save the best looking (in focus) image (findZ.py; Table 2). The Raspberry Pi can run the Python reconstruction software, but a laptop is much faster, has more resources (e.g., memory and hard drive), and provides a more familiar programming environment.

Because of the COVID-19 lockdown of the San Francisco State University campus in Fall 2020, students had to assemble their microscopes at home. Mailing each of the 22 students the set of tools described in the previous paragraph was financially impractical and raised safety issues (e.g., hot soldering iron and sharp tools) because they would not be used under faculty supervision. To compensate, we prebuilt microscope components, performing the necessary drilling and soldering that required tools, and mailed each student a kit with double-stick tape and silicone sealer to finish the assembly and perform the laser-to-image sensor alignment. Assembly consisted of gluing a glass slide between 2 PVC cards with a 3/4-inch (2 cm) center hole that acts as a sample well, gluing it to the top of the predrilled case, then adjusting the cap for the best laser-to-image sensor alignment.

Instead of relying on the apps raspistill and raspivid (69), provided in the Raspberry Pi OS distribution to capture images and video, respectively, which would require a separate USB keyboard, mouse, power supply, and HDMI monitor, we provided instructions on how to install showmewebcam (70), which is an alternative way to collect images and video. Students download an open source disk-image (collection of files) onto an SD card and insert it in the microscope's Pi single-board computer. The disk-image contains the minimum necessary Raspberry Pi OS functionality to run an application that turns the Pi's camera into a USB video device. When plugged into a laptop's USB port, the microscope boots in 3 s and operates as a web camera, allowing any web camera software to be used, including streaming holographic videos over a Zoom call. Additionally, the code on the SD card includes a lightweight application to control camera parameters, including exposure and contrast, essential for producing holographic images with many fringes to produce good reconstructions.

After successfully building their personal microscope, all students used samples of their hair to practice capturing good-quality holograms and perform measurements. Two strands of hair were mounted to a glass microscope slide with clear nail polish in an “X” pattern (Fig 7a). The sharpness of the hair intersection provides a visual guide to select the best sample-to-image sensor reconstruction Z distance. The resulting hologram was reconstructed, and the pixel brightness of a cross section of the hair (Fig 7b) was plotted with ImageJ to determine the hair diameter (Fig 7c).

Once they successfully measured hair diameter, each student developed a project driven by their own questions, in which they measured the microscope's performance, modified microscope hardware or software, or applied the microscope as a tool to answer their research question. In subsequent terms of teaching the course, we allowed students to use other microscopes and image datasets from the public domain, research labs they worked in, or our Center for Cellular Construction partner labs. We shall now share several research projects to demonstrate the variety of work the Optical Engineering Laboratory students conducted.

Two students (J. Luo, Z. Dean) examined the resolution of the holographic microscope and what improvements could be made to increase performance. The project aimed to contribute to a larger class goal of identifying cost-effective approaches to enhance the performance of the microscope. Their hypothesis was that shorter wavelength and use of a pinhole would improve resolution from previous work by Amann et al. (71). In the cited work, the authors built and compared holographic microscopes with a blue laser (405 nm) coupled into a single-mode fiber and a blue LED (430 nm) filtered with a 25-μm pinhole, resulting in a lateral spatial resolution of 1.55 μm and 3.91 μm, respectively, determined with a 1951 USAF resolution test target slide (R1DS1P, Thorlabs, Newton, NJ). With the goal of maintaining the low cost and construction simplicity of our holographic microscope design, the students tested 3 configurations and compared their resolution results to Amann's work.

With a USAF test slide to measure resolution, the students tested a blue laser (405 nm), with and without a 25-μm pinhole (P25k, Thorlabs, Newton, NJ), and the red laser (650 nm) used in the student microscope. They varied the object-to-image sensor distance to achieve the maximum resolution, fixing the laser-to-image sensor to 25 mm. They achieved a 2.76-μm resolution with the blue laser and the pinhole, 4.38-μm resolution with just the blue laser, and 4.92-μm resolution with the red laser (the original student microscope design), verifying their hypothesis (Fig 8). Considering the student microscope did not use a single-mode fiber or pinhole to increase the coherence, both of which would require careful alignment and additional expense, achieving 4.92-μm resolution was an impressive result.

The goal of this project was to automate the loading of samples into the microscope's viewing chamber, for example, to monitor the plankton in a fish tank or from a lake. The student's (D. Ruiz) hypothesis was that a single electric pump under computer control could clear out old samples and load new samples for observation. The system, illustrated in Figure 9, operated as follows: A MOSFET transistor (72) enabled the microscope's Raspberry Pi to control a 12 V peristaltic pump (NKP-12V-S10B, Kamoer Fluid Tech Co., Shanghai, China) (Fig 9b).

The microscope turned on the pump for 30 s to draw water from a plankton reservoir (Fig 9a), delivering a new sample to a 3D-printed flow cell (green box in Fig 9c) while washing out the previous sample into a waste container. After a 5-s pause to allow the water in the flow cell to settle, a 30-s holographic video was recorded. The cycle was repeated often enough to prevent water from drying and fouling the sample well glass, verifying the student's hypothesis. This project was designed and built by a nonengineering biochemistry major and was their first experience with soldering, mechanical design, 3D printing, circuit design, and assembly. The student learned and used Fusion 360 (version V.2.0.6263, Autodesk Inc., San Francisco, CA) to design and 3D print the flow cell (Fig 9c).

The viewing area of the lensless microscope is typically the size of the image sensor (3.76 × 2.74 mm). One student (R. Cisneros) investigated a method to increase the viewing area of the microscope using multiple light sources. The hypothesis was that a horizontal displacement of the light source would expose the edges of the viewing area, and by using multiple light sources under computer control, a mosaic of images could be collected, increasing the field of view. To simplify the design, an array of LEDs manufactured in an epoxy case were used instead of multiple lasers. An 8 × 8 array of red LEDs (Fig 10a), each with a 2-mm-diameter on a 2.5-mm pitch arranged in a plane (KWM-30881CVB, Adafruit, New York, NY), was mounted above the image sensor (Fig 10b,d). By sequentially energizing selected LEDs one at a time and capturing an image, the microscope's effective field of view was enlarged (Fig 10e), verifying the student's hypothesis. The student also demonstrated that increasing the distance between the LEDs and image sensor increased the image contrast (Fig 10c), because a distant LED better approximated a point light source.

The original microscope design had a hard rubber cap (25-mm cube) to hold the red laser directly above the image sensor and block ambient light. The cap was hinged to the case to allow samples to be loaded into the sample well. The hypothesis was that by placing the sample at fixed distances from the sensor, discrete magnifications could be selected. To test this hypothesis the student (C. Guzman) learned Fusion 360 and used the program to design and 3D print a cap with multiple pairs of slots to accept conventional 1 × 3 inch (2.5 × 7.6 cm) microscope slides (Fig 11). Because the size of the interference pattern increases with the distance between the sample and image sensor, multiple pairs of slots provided a set of discrete magnifications. The student tested the cap with a variety of samples, demonstrating discrete magnifications and proving his hypothesis.

The previous project provided slots to select discrete magnifications. Two students (C. Ly, A. Martin) took different approaches to build an apparatus to adjust magnification continuously. The hypothesis was that by continually varying the sample-to-image sensor distance, a continuous zooming feature could be achieved. The first student constructed an apparatus at home (because of COVID-19) with material available around the house (Fig 12a,b). A cardboard frame held a pulley made from a plastic straw and bottle cap. Four strings suspended the image sensor below the fixed sample, continually changing the image-to-sample distance.

The second student built an apparatus at school with a 3D design provided online from the Universidad de Colombia (73) (Fig 12c). The 3D-printed structure used a threaded bolt to move the image sensor above the sample, with the laser fixed below the sample. The student modified the design to accommodate the Raspberry Pi camera and red laser. Both designs verified the hypothesis that a variable zoom could be implemented by mechanisms that moved the sample-to-image sensor distance in a continuous motion.

We provided students image processing code to reconstruct holographic images and detect objects with a simple keyboard interface to change program variables. One class lecture demonstrated how to use TKinter (version 8.6, Python Software Foundation), a Python GUI library to design a mouse-driven GUI. Inspired by this exercise, the student's (S. M. Ahmed) hypothesis was that a GUI would improve the user experience of the reconstruction code. One result was an application (Fig 13, left) that allowed students to use sliders to vary the quantizing threshold, minimum and maximum area of detected objects, and label and save the objects. Another application (Fig 13, right) enabled students to use sliders to view their holographic videos frame-by-frame and modify detection parameters. The student's GUIs were adopted by the class as the preferred method of processing holographic images, proving the hypothesis.

Two students (E. Samperio, A. Barrera-Velasquez) teamed up to classify plankton. One student created a labeled training set of 11 plankton classes (131020, Carolina Biological Supply, Burlington, NC) imaged by the entire class with their holographic microscopes by cropping and reconstructing 360 images from videos recorded by all class members. The other student performed object segmentation (Fig 14a–d), feature extraction, and k-means clustering (Scikit-learn) by modifying code from the image processing pipeline.

The hypothesis was that the plankton classes could be classified by area, aspect ratio, and circularity. The segmentation process separated the object (Fig 14a) from the background by blurring the image (Fig 14b) to remove object details, which might otherwise make one object look like several smaller objects. A quantization threshold was applied to the blurred image and the OpenCV contour function determined the object contour (Fig 14c,d). The contours were used to calculate the geometric features of area, aspect ratio, and circularity. These 3 features were used to cluster the objects into 11 classes (Fig 14e) by the k-means clustering algorithm. Although the results did show that the classes presented sensitivity to the features, the combination of the 3 features were not sufficient for classification, disproving the hypothesis.

A group of students (A. Nelson, J. Suarez, G. Alarcon-Cruz) concerned about the dangers of microfibers, primarily from clothing, in waterways designed a method to distinguish cotton from synthetic microfibers. Their hypothesis was that the holographic microscope could be used as a low-cost tool to monitor microplastic pollution in the environment. They collected lint from several clothing dryers, separated out the tiny fibers, suspended them in a dilute solution with water, then captured images with their holographic microscope. They used ImageJ to measure the path and endpoint lengths, and used the ratio to distinguish the 2 types of microfibers (Fig 15). They determined that cotton fibers had greater curvature than the synthetic fibers and can be observed with the low-cost holographic microscope, proving their hypothesis.

Our objectives in designing and teaching the Optical Engineering Laboratory course are to introduce biology and biochemistry students to image processing, instrumentation, and research methods and skills. The software side has 3 levels of achievement: knowing the basics of how to program in Python, learning how to use OpenCV to perform image processing tasks, and composing a program that solves a problem based on this acquired knowledge. The hardware side has similar levels of achievement: getting to know how to use basic tools, applying them to build the microscope, and using them to conduct original research. Most importantly, to conduct research, students need to learn how to define a project that is interesting scientifically and challenging but narrow enough in scope to be completed in one semester. Lastly, communication and collaboration skills are essential for future scientific careers. Students learn how to present results in a variety of formats and collaborate with faculty, fellow students, and researchers outside the classroom. In this section we discuss our impressions of the effect of the course on these goals and share students' impressions, based on their responses to pre- and post-class surveys.

For the Optical Engineering Laboratory course, we were pleasantly surprised how fast students could construct the microscopes and how much they enjoyed doing hands-on work. For many, it was the first time they used a soldering iron and Dremel tool. We received encouraging comments like, “I never knew I liked engineering,” “I really liked soldering,” and “I never knew I could do engineering.” We were particularly impressed by the ingenuity and resourcefulness of the student who built an apparatus for continuously adjusting magnification out of household items. This mirrors our own experience of adapting existing equipment and materials to perform novel experiments in our professional research.

The big challenge in the lab was maintaining student's progress in their research projects. To address this, each week students would stand before the class and give a progress report. We began taking on the role of project managers, identifying critical-path tasks and experiments and setting these as deliverable goals for the next week. We also had to teach students how to present novel research in oral and written presentations. This task is more difficult than writing up the results of an assigned lab experiment, in which procedures are defined and the results are well known. To address this challenge, we devoted a class session in which each student would read and review a research paper published on bioRχiv (74). This activity helped students see how figures, tables, and other parts of a paper must be clear to the reader and provide all the essential elements of the research, so the results can be reproduced by others.

The holographic microscope provides a platform for hands-on learning of several scientific and engineering disciplines. These include physics, mathematics, electrical and mechanical engineering, signal processing, and computer science. In academia, these disciplines are often taught independently in separate departments and different majors. The holographic microscope provides an affordable, accessible, and extendable platform to study and apply all these disciplines in one project-based interdisciplinary class. For many of the biology and biochemistry majors this was their first exposure to engineering and hands-on equipment building, and many were surprised to find how much they enjoyed engineering.

The image processing pipeline provides a flexible structure for inclusion in a syllabus. Each pipeline component can be taught individually because they had defined inputs and outputs and were used in sequence to perform a complex function (i.e., classify objects). This modular feature made it possible to select a few components and integrate them into the Optical Engineering Laboratory course. Many students used individual components in their research projects. Providing working code demonstrated how to take a complex program and break it into functions. The modules taught Python programming skills and introduced students to the power of the OpenCV computer vision library and Scikit-learn machine learning library. An ideal 1-yr curriculum would be a semester-long Image Processing lecture course followed by the Optical Engineering Laboratory course. This combination would provide enough time for students first to learn the theory and gain coding experience with the entire pipeline in the Image Processing lecture course, followed by learning and applying engineering and research skills in the Optical Engineering Laboratory course.

In research, especially in biology and biophysics, we are often faced with buying expensive equipment, or more often modifying and adapting equipment we have, and sometimes building it from scratch. Research labs need graduates who are versed in more than one field, who have depth (the traditional product of education) but also breadth, basic knowledge, and skills in complementary disciplines. The multidisciplinary laboratory and project-based course we developed, supported by lessons in programming and image processing, provided opportunities for students to broaden their knowledge, experience, abilities, and opportunities.

TZ, RE, and Y-HMC conceived and designed the courses. TZ and RE drafted the manuscript. TZ, RE, Y-HMC, NA, AA, JL, and SC edited the manuscript. TZ, RE, Y-HMC, AK, NA, SA, and AA were course instructors. JL and ZD performed microscope performance analysis. TZ and SA wrote code for the course. MP performed student evaluations. SB and SC provided project oversight. The student consortium comprised Gilbert Alarcon-Cruz, Adrian Barrera-Velasquez, Rocio Cisneros, Carlos Guzman, Cara Ly, Adrian Martin, Alicia Nelson, Donovan Ruiz, Emily Samperio, and Jessenia Suarez.

This material is based on work supported by the National Science Foundation (DBI-1548297). We thank the students listed in the Acknowledgments for their participation, hard work, and enthusiasm to learn and discover. Disclaimer: Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. The authors declare no competing interests.