Machine learning: The future of healthcare

Aug. 18, 2021
There is no doubt that the future of photonics in healthcare, especially in diagnostic healthcare, will hinge on machine learning, but it is critical to remember that it's a tool, not a magic wand.

Over the last 12 months, we have featured several photonic techniques poised to revolutionize diagnostic healthcare: Raman spectroscopy, multiphoton fluorescence microscopy, and most recently, discrete frequency infrared (DFIR) microscopy. We have discussed how these technologies could provide clinicians with faster and more accurate information, enabling better decisions about a patient’s care plan. However, we have not spent much time discussing the challenges associated with translating these technologies from laboratory to clinical settings. In the world of translational medicine, this is known as the “valley of death” since so few technologies manage to cross the preclinical-to-clinical divide. Biophotonics is no exception. In fact, Bruce Tromberg, director of the Beckman Laser Institute and Medical Clinic at the University of California Irvine (Irvine, CA), noted during a presentation on entrepreneurship in healthcare photonics at the 2020 BiOS Photonics West conference, “There isn’t one valley of death. There are many.”1   

Historically speaking, one of the biggest challenges to clinical translation for advanced optical techniques is the sheer quantity of data needed to ensure specificity over the broader population. Implementation on scale has proved to be nothing short of a massive undertaking. To illustrate this point, let us look at the scenario of a researcher using Raman spectroscopy to differentiate between cancerous and benign tissue samples.

To translate this technology from the lab to the clinic requires samples to be collected from hundreds, if not thousands, of patients across a wide range of demographics. These samples would then be measured with the Raman spectrometer and cross-validated and labeled by a certified pathologist, then preprocessed and fed into a multivariate classification algorithm. For years, both the pathology and computational requirements seemed insurmountable. Luckily, over the past few years, developments in machine learning and a shift in perception within the pathology community toward automation2 appear to be bridging those valleys.

“Having applied multiple techniques over several health-based machine learning applications, the main issue by far is the access to rich applicable datasets. Data that human-labeled is typically small, and with human error rates on some medical digital data accuracy, well under 90%, that too can be a limitation,” says John Murphy, CEO of Stream.ML and Bio-Stream Diagnostics (Edmonton, AB, Canada). “Fortunately, newly deep learning techniques that work with smaller datasets, that are ultra-accurate, means that AI is overcoming these classic roadblocks.” 

Fundamentals of machine learning

Since it appears that machine learning may genuinely be the key to the future of photonics in healthcare, it is vital first to understand what machine learning is before delving into specific examples from the published literature. On the most fundamental level, machine learning is a catch-all term for any set of algorithms designed to take in data, learn from it, and use that knowledge to detect patterns.

Obviously, that statement could not be more allied with the goal of automated imaging, spectral, and hyperspectral diagnostic systems. Still, it fails to provide an intuitive understanding—a look at the wizard behind the curtain. The only way to truly grasp how machine learning aids photonic diagnostics is to understand the underlying algorithms. 

Loosely based on biological systems, neural networks form the foundation of machine learning technology. But unlike many would be led to believe, you do not need to be an expert in biology or computer science to gain a fundamental understanding of neural networks. In fact, any optical engineer who has ever optimized a merit function has already done much of the heavy lifting involved in building a neural network without even knowing it.

The fundamental principle that underlies every neural network is the concept of neuron activation. 

Again, this sounds complicated, but it really is not. The easiest way to think about a neuron is as a matrix element, and its activation is the element’s value. For example, in the neural network shown in Figure 1, each neuron in the input layer corresponds to the pixel number in the spectrum. Thus, the normalized pixel value is the input activation. The output of each input layer neuron is then fed into each of the neurons in a second layer. The exact details of the mathematics of the output function are not that important to understand. What is essential to understand is that the activation of each neuron in a layer is determined by the weighted sum of the output from all of the neurons in the previous layer. 

If we assume that the input spectrum (see Fig. 1) contains 4000 neurons (pixels) and the next layer is comprised of four neurons, to compute the final activation of the neurons in the second layer, you would need to assign a total of 16,000 weights. That would be impossible to perform manually, and we have not yet gotten to the third layer. That is where the actual machine learning comes into play.

In a neural network, all the weights start completely random, and instead, you only focus your attention on the output layer through a process known as training. For training, what you do is input a known spectrum as well as a known output—for example, cancerous or benign. This is accomplished by assigning one of the two output nodes as cancerous output and the other as benign. Then, using a least-squares merit function, known as a cost function in machine learning, and a gradient descent algorithm to optimize each individual weight (via backpropagation), the system can learn which weights are needed to ensure that each time a cancerous spectrum is inputted, the proper output neuron activates, and vice versa.

This largely blind process is where the actual machine learning occurs since it involves no human involvement other than feeding the network a large enough human-labeled dataset. For this reason, all the intermediary layers in the network are called hidden layers. It is important to note that this is a highly simplified example of a neural network and is not even the most used architecture these days.

Convolutional neural networks are currently the most popular for these types of applications. Still, it serves as an excellent visual explanation of how neural networks and machine learning works and how powerful this platform can be for automating spectral and image classification. 

It is also important to point out that with the TensorFlow open-source machine learning platform release in 2015, researchers no longer need to worry about the complex mathematics underlying neural networks. Today, it is possible to build a neural network capable of image classification in roughly 15 lines of code. With this much computation power so easily implementable, it is no surprise that machine learning is showing up in all sorts of applications, including spectral and image classification. 

Machine learning in spectral diagnostics

One fascinating example of how machine learning is used with spectral data comes from a highly interdisciplinary team from Stratford University in Virginia. The team was able to accurately identify 30 different bacterial pathogens using machine learning to surface-enhanced Raman spectroscopy (SERS).

In a paper titled “Rapid identification of pathogenic bacteria using Raman spectroscopy and deep learning,” published in Nature Communications, described how they were able to use a convolutional neural network (CNN) with 25 layers to determine the bacterial class and identify antibiotic treatment with a 97% accuracy (see Fig. 2).3 All of the spectra were collected using a Horiba (Piscataway, NJ) LabRAM HR Evolution Raman microscope with a 633 nm excitation laser and an integration time of only 1 second. A simple visual inspection of the 30 different bacterial spectra shown in Figure 2 underscores the impressiveness of this feat when you consider both how noisy the signal is (typical SNR of 4.1) and how similar the spectra are to each other.

FIGURE 2. Visualization of the process of (a) preparing the bacterial samples on a gold SERS substrate, (b) collecting the spectra, and (c) inputting the spectra into a convolutional neural network to determine the bacterial class and antibiotic treatment, as well as sample spectra for all 30 bacterial samples (d) [3]. 

According to the study authors, “bacterial infections are a leading cause of death in both developed and developing nations, taking >6.7 million lives each year.” Therefore, quickly and accurately identifying and executing the appropriate antibiotic treatment could save a large percentage of those lives every year. Furthermore, even though this research was conducted using a laboratory-grade Raman microscope, there is no reason this technology could not be developed into a point-of-care screening tool. 

Another exhilarating prospect for spectroscopic diagnostics is for the diagnosis of Alzheimer’s disease. Traditionally, this requires either expensive brain imaging or dangerous cerebrospinal fluid testing. However, a team led by Noureddine Melikechi, dean of the Kennedy College of Sciences at the University of Massachusetts Lowell, has shown promising results for identifying Alzheimer’s in plasma using a combination of laser-induced breakdown spectroscopy (LIBS) and machine learning.4 With 44 million people currently living with Alzheimer’s worldwide, the ability to detect biomarkers for the disease with a low-cost, minimally invasive procedure would be truly revolutionary.

Additionally, researchers at the Fujian Normal University, Fujian Medical University Union Hospital (Fujian, China), and Yachay Tech University (San Miguel de Urcuquí, Ecuador), which has successfully used convolutional neural networks and multiphoton microscopy, classify the severity of hepatocellular carcinoma (liver disease) automatically.5  

Jianziong Zhu and Zhihao Ren from the National University of Singapore also demonstrated the feasibility of plasma-enhanced IR absorption spectroscopy to detect volatile organic compounds in a breath using machine learning.6

It is important that researchers not become overzealous in the application of machine learning tools at the expense of good science. In a recent article published in IEEE Spectrum, Megan Scudellari rightfully reported that the lack of reproducibility in publications related to machine learning in health research “faces a reckoning” due to the lack of reproducibility.7 In that article, she wrote, “healthcare is an especially challenging area for machine learning research because many datasets are restricted due to health privacy concerns and even experts may disagree on a diagnosis for a scan or patient. Still, researchers are optimistic that the field can do better.”

REFERENCES

1. See https://bit.ly/HealthcarePhotonics.

2. S. Rattenbury, The Pathologist, 3, 5, 30–33 (2016). 

3. C. S. Ho et al., Nat. Commun., 10, 1, 1–8 (2019); doi.org/10.1038/s41467-019-12898-9.

4. R. Gaudiuso et al., Spectrochim. Acta. Part B At. Spectrosc., 171, 105931 (2020); doi.org/10.1016/j.sab.2020.105931.

5. H. Lin et al., J. Biophoton., 12, 7, p.e201800435 (2019); doi.org/10.1002/jbio.201800435.

6. J. Zhu et al., ACS Nano, 15, 1, 894–903 (2020); doi.org/10.1021/acsnano.0c07464.

7. A. Mannini et al., Sensors, 10, 2, 1154–1175 (2010); doi.org/10.3390/s100201154.

About the Author

Robert V. Chimenti | Director, RVC Photonics LLC

Robert V. Chimenti is the Director of RVC Photonics LLC (Pitman, NJ), as well as a Visiting Assistant Professor in the Department of Physics and Astronomy at Rowan University (Glassboro, NJ). He has earned undergraduate degrees in physics, photonics, and business administration, as well as an M.S. in Electro-Optics from the University of Dayton. Over a nearly 20-year career in optics and photonics, he has primarily focused on the development of new laser and spectroscopy applications, with a heavy emphasis on vibrational spectroscopy. He is also very heavily involved in the Federation of Analytical Chemistry and Spectroscopy Societies (FACSS), where he has served for several years as the Workshops Chair for the annual SciX conference and will be taking over as General Chair for the 2021 SciX conference.

Sponsored Recommendations

How to Tune Servo Systems: Force Control

Oct. 23, 2024
Tuning the servo system to meet or exceed the performance specification can be a troubling task, join our webinar to learn to optimize performance.

Laser Machining: Dynamic Error Reduction via Galvo Compensation

Oct. 23, 2024
A common misconception is that high throughput implies higher speeds, but the real factor that impacts throughput is higher accelerations. Read more here!

Boost Productivity and Process Quality in High-Performance Laser Processing

Oct. 23, 2024
Read a discussion about developments in high-dynamic laser processing that improve process throughput and part quality.

Precision Automation Technologies that Minimize Laser Cut Hypotube Manufacturing Risk

Oct. 23, 2024
In this webinar, you will discover the precision automation technologies essential for manufacturing high-quality laser-cut hypotubes. Learn key processes, techniques, and best...

Voice your opinion!

To join the conversation, and become an exclusive member of Laser Focus World, create an account today!