AI over Thermal Images for Medical Applications: Introduction to Thermal Imaging

Why Thermal Imaging Is Essential for Advancing Healthcare? 

In many developing and underdeveloped countries, providing healthcare services to rural areas still remains a formidable challenge. At the heart of this challenge lies the need for screening technologies that are not only cost-effective but also non-invasive, portable, and require minimal human expertise. Regrettably, many conventional imaging modalities come with exorbitant costs, radiation risks, and a demand for specialized skills.  This makes the combination of Artificial Intelligence (AI) and thermal imaging a game changer.

Thermal imaging presents a multitude of benefits: it does not involve any external radiation, offers a no-touch, non-contact experience, boasts portability, and remains remarkably cost-effective. During the COVID-19 pandemic, we all have seen their use extensively in airports, malls, railway stations and other public locations for COVID-19 screening. Apart from this, its usage has been studied for breast cancer screening, diabetic foot, muscular skeletal disorders, pain management, thyroid cancer, oral cancer, skin cancer etc.  Its combination with AI also holds the promise of reducing the dependence on scarce human expertise, a critical factor in countries like India, where the ratio stands at just one radiologist for every 100,000 people.

In the inaugural series of our blogs, we embark on a journey to unravel the fundamental concepts of thermal imaging. These insights serve as stepping stones for exploring the transformative potential of AI in the realm of medical applications using thermal imaging.

What is a Thermal Image?

Long wave infrared radiation (LWIR) is naturally emitted by every object, including the human body. Infrared thermal cameras convert this LWIR emitted from an object (or a body) and convert them into a temperature matrix. Each element in the temperature matrix represents a temperature value corresponding to a particular point on the object's surface.

                                      

The size and area of the object that is being imaged is determined by the camera's spatial resolution, Field of View (FOV) and its positioning from the object. While the spatial resolution of a thermal camera influences the level of detail captured in the temperature matrix, FOV and camera position decides the area of object. Higher spatial resolution allows for finer granularity and more precise temperature measurements, while lower resolution may result in a coarser representation of the temperature distribution.

The latest advanced thermal cameras have spatial resolution of 1024 × 768 pixels or more with a temperature sensitivity and accuracy error of 20 mK (0.02°C) and ± 1°C .  This means that we can measure minute relative temperature differences of up to 0.02°C with a maximum deviation of ± 1°C.

To reach to this level of sophistication, thermal cameras have advanced over the years. If you are interested to know more about the advances in thermal cameras over years,  you can read the paper -  Kakileti ST, Manjunath G, Madhu H, Ramprakash HV. Advances in breast thermography. New Perspectives in Breast Imaging. Malik A (ed): IntechOpen, London. 2017 Oct 4:91-103.

                                             

In the context of a human body, the obtained temperature values are real numbers and typically lie between 20℃  and  40℃. Given the continuous range of temperature values, directly studying and analyzing the numerical data with the naked eye becomes impractical. To facilitate visual interpretation and analysis, a mapping function is employed to convert the temperature values into color images a.k.a thermal images that are easier to comprehend. These mapping functions are achieved through the use of color palettes that typically present on the infrared thermal image.

Color Palettes

Color palettes are a predefined sequence of colors that are stacked vertically, forming a visually appealing gradient. The purpose of these color palettes is to establish a correspondence between temperature values and colors, enabling the visualization of temperature variations across the object's surface. To generate a color thermal image from the temperature matrix, a minimum temperature (Tmin) and a maximum temperature (Tmax) are selected as the limits for the color mapping. The temperature range between Tmin and Tmax is divided into multiple bins or intervals, with the number of bins corresponding to the number of colors present in the chosen color palette. Each temperature bin is then mapped to a specific color value in the color palette, resulting in a color image that represents the temperature distribution across the object.  Some prominently used color palettes are described below.

Rainbow Palette

Rainbow color palette is the most commonly used color palette for interpretation of medical thermal images. As the name says, it draws the color shades of the rainbow to form the color palette. An inherent advantage of this palette lies in its ability to vividly depict elevated temperature zones in striking red hues, while rendering cooler temperatures in a gentle interplay of soothing tones such as green and blue. This deliberate arrangement not only lends visual appeal but also confers clarity and facilitates the identification of distinct thermal patterns.

                                           

Grayscale Palette

Grayscale palette converts the temperature matrix into a gray scale image by mapping the temperature values into a numerical range of 0 to 255, which represents the intensity or brightness of the gray color. Variation in the shades of gray allows one to study the variations in the heat patterns. Cooler temperature regions appear darker or closer to black, while warmer temperature regions appear lighter or closer to white. This grayscale representation allows for a straightforward visualization of temperature differences without the distraction of color information. Grayscale palettes are commonly used for visualizing structural details, such as vessels, bones or tissue boundaries.

                                                

                                      

Inverted Grayscale Palette

Inverted grayscale palette is an inverse of gray palette formed by stacking the grayscale palette in reverse order. Unlike in the grayscale palette, cooler temperature regions appear lighter or closer to white and warmer temperature regions appear darker or closer to black in inverted grayscale palette. Inverted grayscale images are often of interest in identifying areas of inflammation, blood flow irregularities, and detecting certain medical conditions.

                                            

Iron Hot Palette

Iron Hot color palette is a commonly used color palette for visualizing thermal images for industrial applications. However, its usage in medical imaging is very limited. This distinctive palette is characterized by an array of resplendent yellow hues, with the zenith of heat represented by the pristine purity of white.                        

    

Camera Parameters for Accurate Temperature Values

As you might have guessed, temperature values are crux for the interpretation and generation of thermal images. However, thermal camera assumes default values for several parameters in order to obtain the temperature values from the emitted long wave radiation. To understand more, first, we need to understand how thermal cameras obtain temperature profile. When a thermal camera is positioned to measure the temperature profile of a target object, it not only receives the emission of the target object (Eobj) but also  receives two other emissions: emission of surrounding heat that is reflected off the object (Erefl) and emission of atmosphere (Eatm). Refer the below figure for more clarity. In order to obtain the temperature profile of the target from the total emitted long wave radiation received by the thermal camera (Wtot), thermal cameras assume default values for several parameters as shown in the below equation.  It is important that one needs to set the below parameters to their true values in order to obtain accurate temperature values.

Emissivity (ε): Emissivity of an object is defined as its capability to radiate thermal infrared radiation when compared to a blackbody at the same temperature. Emissivity ranges from 0 (Perfect reflector such as mirrors) to 1 (Perfect absorber such as blackbody). The emissivity of the human skin surface is considered to be approximately 0.98. Therefore, for the purpose of medical thermography, an emissivity of 0.98 is commonly used. However, many cameras come with a default emissivity of 0.95. So it is important to modify this value and many camera vendors such as FLIR allow these values to be changed even after the image capture.

Reflected Temperature (Tref): Reflected Temperature corresponds to the amount of thermal infrared radiation emitted from background or surrounding objects that is reflected off the measuring object. This is one of the reason for the importance of ensuring no strong heat sources or reflective surfaces near the imaging area. Further, the level of influence of Tref on the actual measured temperature depends on the emissivity of the object. For objects with very high emissivity, this influence is minimal, however, this influence is very high for objects with low emissivity. Assuming there are no strong heat sources in the imaging area, this value can be set to environmental temperature ( around 68°F/20°C) for the purpose of medical thermography.

 Atmospheric Transmissivity (𝜏atm):  This is defined as the atmospheric ability to transmit the thermal infrared radiation. In the context of medical thermography, atmospheric transmissivity is the ability of the air to transmit thermal infrared radiation. 𝜏atm depends on the relative humidity and the distance between the measuring object and the camera.  In most scenarios,  estimated 𝜏atm will be very close to 1, thereby, having a minimal influence on the temperatures.

Atmospheric Temperature (Tatm):  Tatm refers to the temperature of the atmosphere where the imaging is being conducted. This can be simply obtained by using standard room thermometers. This parameter also have minimal influence as (1-𝜏atm ) tends to 0 in many scenarios for medical thermography.

Boltzmann Constant (σ): This is fixed value of 5.67 × 10-8W m-2 K-4 and there is no need for any adjustment.

Let us see a demo of how these parameters impact the temperature values

As can be seen from the above demo, the temperature value at sp1 was 29.5°C with default values of emissivity, Tref , Tatm and 𝜏atm . However,  when these parameters were changed, it resulted in variation of measured temperature values. It is also important to note that this variation might be uniform or constant variation across the pixels in the image.

Even though you adjust the above parameters, it is also important that one need to regularly calibrate thermal cameras for accurate temperature measurements. This is necessary because over time, electronic components within a thermal camera may age, deteriorate, or undergo changes. These changes can lead to a gradual drift, known as calibration shift, in the camera's temperature readings, resulting in inaccurate temperature measurements. Calibration is essential to correct for such drift and ensure that the camera continues to provide accurate measurements. The frequency of calibration may vary based on factors such as camera usage, environmental conditions, and the required level of accuracy for specific applications.

Find out more on the challenges with manual interpretation and latest research on AI over thermal images for various medical applications in the next blogs.    

Do post your questions in the comments section!