Diagnostic Performance of Thermal Imaging of Oral Cancer
Effectiveness of a non-radiating, noninvasive digital infrared thermal imaging system in the detection of cervical lymph node metastasis from oral cavity cancer
Fan Dong1,2,3,4,5,*, Chuansibo Tao6,*, Ji Wu7, Ying Su7, Yuguang Wang1,2,3,4,5, Yong Wang1,2,3,4,5, Chuanbin Guo3,6 and Peijun Lyu1,2,3,4,5
1National Engineering Laboratory for Digital and Material Technology of Stomatology, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China.
2Center of Digital Dentistry, Peking University School and Hospital of Stomatology, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China.
3Department of Prosthodontics, Peking University School and Hospital of Stomatology, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China.
4Research Center of Engineering and Technology for Digital Dentistry of Ministry of Health, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China. 5Beijing Key Laboratory of Digital Stomatology, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China.
6Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology, 22 Zhongguancun Avenue South, Haidian District, Beijing 100081, PR China.
7Tsinghua-Rohm Electronic Engineering Hall 8-301,Tsinghua University, Beijing, 100084, PR China.
* These authors have contributed equally to this work
Correspondence to: Chuanbin Guo, email:
               Peijun Lyu, email:
Key words: digital infrared thermal imaging; cervical lymph node metastasis; thermography
Abstract
The objective of this study was to evaluate the diagnostic performance of a non-radiating, noninvasive infrared (IR) thermal imaging system in the detection of cervical lymph node metastasis from oral cavity cancer. We performed the first application of digital IR thermal imaging to detect cervical lymph node metastasis from oral cavity cancer. In this prospective clinical trial, a total of 90 oral cavity cancer patients suspected of having cervical lymph node metastasis underwent IR imaging of the neck prior to neck dissection. Analysis of the IR images was performed by two different methods: manual qualitative analysis and automatic analysis by an entropy-gradient support vector machine (EGSVM). Compared with manual qualitative analysis, the EGSVM-based automatic analysis had higher sensitivity (84.8% vs. 71.7%), specificity (77.3% vs. 72.7%), accuracy (81.1% vs. 72.2%), positive predictive value (79.6% vs. 73.3%) and negative predictive value (82.9% vs. 71.1%). Therefore, this digital IR imaging system is a promising non-radiating, noninvasive tool for the detection of cervical lymph node metastasis from oral cavity cancer.
Introduction
Oral cavity cancer (ICD-10:C00-C08) is a serious and growing problem in numerous countries. Worldwide, in 2012, there were approximately 300,400 new cases and 145,400 deaths from oral cavity cancer[]. Controlled by the primary tumor, the metastasis to the cervical lymph node is the most significant factor that determines prognosis[]. Regardless of the site of the primary tumor, the 5-year survival rate drops by nearly 50% if a single metastatic lymph node is present in either side of the neck. The survival rate of patients with a single metastatic lymph node in both sides is reduced to only 25% of that in patients without lymph node metastasis[]. Therefore, assessment of lymph node involvement is of utmost importance in patients with oral cavity cancer.
The current diagnostic modalities for the detection of lymph node metastasis include computed tomography (CT), positron emission tomography/computed tomography (PET/CT), magnetic resonance imaging (MRI), ultrasound, and ultrasound-guided fine needle aspiration cytology (FNAC). Each technique has its own unique advantages and disadvantages. In clinical practice, we must compromise between accuracy and effectiveness vs. invasiveness and cost considerations. An innocuous, noninvasive imaging modality remains an open quest in biomedical imaging. Among the current modalities mentioned above, only MRI and ultrasound do not involve radiation exposure or invasive procedures. Although ultrasound is a widely available imaging modality, its diagnostic performance is highly dependent on the experience of the operator[]. With a relatively expensive cost, MRI is not widely available. Therefore, a reliable, new, cost-effective method is needed.
Infrared (IR) thermal imaging is a non-radiating, noncontact, noninvasive, low-cost and fast imaging modality that passively captures thermal radiation emitted by any object above absolute zero. Unlike other imaging modalities, IR imaging is for functional rather than anatomical information. Temperature is a useful indicator of diseases. Previous studies have found that temperature distribution is symmetrical between the two sides of the human body in healthy people; in diseased individuals, abnormal blood flow results in abnormal temperature distribution[]. For instance, excess heat generated by blood flow (angiogenesis) and metabolic activity in breast cancer provide the basis for the detection of breast cancer with IR imaging[]. With technology advances in thermal cameras and image analysis tools over the years, there has been a resurgence in the use of IR thermal imaging as a diagnostic tool in medicine[]. IR imaging has been applied to the diagnosis of many diseases such as breast cancer[, ], melanoma[], diabetes[], infantile hemangiomas[] and lower extremity deep venous thrombosis[]. To the best of our knowledge, IR imaging has not been applied to the detection of cervical lymph node metastasis.
In this study, we proposed a digital infrared thermal imaging system as a screening and diagnostic tool for the detection of cervical lymph node metastasis from oral cavity cancer. Analysis of the IR images was performed by two methods: manual qualitative analysis and automatic analysis by an entropy-gradient support vector machine (EGSVM). We also investigated the diagnostic performance.
Methods
Study design
This prospective study was approved by the Bioethics Committee of Stomatology Hospital of Peking University, Beijing, China. (NO. PKUSSIRB-201628047), and all patients signed informed consents prior to entering the study. The study protocol is shown in Figure 1. We enrolled a series of 90 patients (60 male [66.7%], 30 female [33.3%]) ranging from 29 to 81 years (mean=58.2 years, SD=12.3 years) who were scheduled for neck dissection with resection of previously untreated primary oral cancer. Of the 90 patients in our study, 55 (61.1%) were clinically N0 based on physical exam, while 35 (38.9%) were clinically N+. The site and histological type of the primary tumor are provided in Table 1. Apart from the diagnostic biopsy of the primary tumor and previous routine dental treatment, no patients had undergone previous head and neck surgery, chemotherapy, or radiotherapy.
IR examination
IR imaging of the neck was performed 1 day before surgery. IR examination was performed by one radiological technician with a thermographic system (Avio R500 Thermal Imaging System, NEC Corporation, Japan), which was an uncooled micro-bolometer with a focal plane array detector. The image matrix size was 640-480, with a response wavelength of 8-14 μm, and a temperature resolution <0.025°C. The procedure was performed in a temperature-controlled room maintained between 23°C and 25°C and 50% relative humidity. Each participant was asked to sit on a chair in an erect position, with the neck exposed, at a distance of approximately 0.5 meters away from the IR camera. After 15 minutes of rest, IR images of the frontal neck were taken.
Manual qualitative analysis of IR imaging
Qualitative analysis was performed by two experienced head and neck radiologists who were unaware of the histological results using Infrec Analyzer 2.6 software (NEC Corp., Japan) with manual brightness and contrast adjustment. Disagreements between two radiologists were resolved via consensus. In this study, IR criteria for the detection of metastasis were modified from those used in breast cancer [, ]. The presence of at least one of the following criteria (Figure 2) was considered a positive indicator for cervical lymph nodal metastasis: (a) increased vascular density with a tortuous vascular morphologic pattern or aberrant vasculature in the region of interest (ROI) but not in the contralateral side; (b) unilateral dilated vasculature such as a facial artery, a submental artery or a carotid artery; (c) a surface temperature difference >1°C in the ROI compared to the mirror image site on the contralateral neck; (d) a bulging outline contour with elevated surface temperature in and around the ROI.
Surgery and histopathological examination
The range of neck dissection was based on the criteria of the Stomatology Hospital of Peking University. The indications and choice of a neck dissection should be determined based on preoperative examination results and intraoperative findings. Patients with evidence of clinical N2 or N3 would undergo a radical neck dissection (RND) whose scope of the surgery involves cervical lymph nodes level I, II, III, IV and V. The preservation of important anatomical structures such as the internal jugular vein, the accessory nerve, the sternocleidomastoid muscle, etc. depends on the relationship between these anatomical structures and the suspected metastases observed during the surgery. Patients with evidence of clinical N1 should undergo a selective neck dissection such as supraomohyoid neck dissection (SOHND) or extended supraomohyoid neck dissection, whose scope of surgery involves respectively cervical lymph nodes level I-III or level I-IV. For malignancies with high risk of cervical metastasis, a selective neck dissection is required even if the preoperative impression is cN0. The contralateral neck of a metastatic lymph node should undergo a selective neck dissection or functional neck dissection considering the risk of contralateral metastasis. Additionally, if the primary tumor crosses the midline, a bilateral neck dissection is also necessary[].
At the time of surgery, the partitioned surgical neck specimens were separated by surgeons and fixed in 10% buffered formalin. The dissected lymph nodes were processed and stained with hematoxylin and eosin for pathological assessment. A routine pathological evaluation of the lymph nodes was performed by two pathologists on one or two sections[]. Diagnoses were reviewed by one pathologist with 10 years of experience. All nodes were recorded as positive or negative for metastasis.
Automatic analysis by EGSVM
EGSVM is mainly a four-step procedure. In the first one, the region of human neck in raw images are cropped and converted to grayscale images. An original IR image and the corresponding cropped gray image are shown in Figure 3. Second, features used for classification are extracted from cropped gray images. Third, model parameters of a SVM [] classifier are trained on the features obtained from second part. Finally, the model is used in the automatic analysis of lymph node involvement.
Feature extraction. During manual qualitative analysis, we found that an asymmetric thermographic pattern, including elevated surface temperature and abnormal vascular pattern, was an important indication of nodal involvement (Table 2). Hence, features which can describe irregularity of an IR image would be helpful to distinguish metastatic lymph images from the rest. During image processing, entropy is a commonly used measure of information contained in an image [, ]. In the feature extraction process, an original gray image is firstly normalized according to different window size as follows:
where win represents a window in the original image. 
means the gray value at point (x, y) of the normalized image, which is obtained by scaling the gray value of corresponding point
in the original image.
If we change the size of window, we can get a series of normalized images with different size, denoted as
, where n represents the quantity of all normalized images. The extending process of win is shown in Figure 4. In our experiment settings, the win extends from central to around evenly.
Furthermore, to emphasize the details of normalized images, we acquire histogram equalized image of each of them, and denoted as 
. For each image Re_imagek, we can calculate corresponding entropy hk as
, 
and these entropies can make up for a vector 
. We can further get the entropy-gradient feature 
based on 
, as according to:
The entropy-gradient feature 
is extracted from histogram equalized image sets in the same way as of the set of normalized images. 
is finally used as the feature vector for the input of SVM classifier.
Model training and predicting. As have been mentioned, we have 90 samples in total. To make better use of these data, nine fold cross-validation method have been applied in the testing. The dataset is randomly divided into nine subsets, each containing equal number of samples. The nine subsets are then grouped into a training set and a testing set. The training set consists of eight of these subsets and the testing set consists of the remaining one. This procedure is repeated nine times and every subset is used once for testing. The final matrices are the average of the five testing results, including accuracy, sensitivity and positive predictive value (PPV) and negative predictive value (NPV).
Results
In our study, a total of 90 patients with histopathological correlations were biopsied. Lymph node metastasis was present in 46 (51.1%) patients. A total of 44 (48.9%) patients did not contain histological evidence of lymph node metastasis. Within the 35 clinically N+ patients, 25 (71.4%) contained histological evidence of lymph node metastasis and 10 (28.6%) did not. Within the 55 clinically N0 patients, 21 (38.2%) contained histological evidence of lymph node metastasis and 34 (61.8%) did not.
Based on IR criteria for manual qualitative analysis, the distribution of lymph node metastases is shown in Table 2. The sensitivity, specificity, accuracy, PPV and NPV by manual qualitative analysis for the detection of lymph node metastasis were 71.7% (95% CI: 58.7%, 84.8%), 72.7% (95% CI: 59.6%, 85.9%), 72.2% (95% CI: 63.0%, 81.5%), 73.3% (95% CI: 60.4%, 86.3%), and 71.1% (95% CI: 57.9%, 84.4%), respectively (Table 3). Compared with manual qualitative analysis, EGSVM-based automatic analysis had an apparently higher sensitivity (84.8% vs. 71.7%), specificity (77.3% vs. 72.7%), accuracy (81.1% vs. 72.2%), PPV (79.6% vs. 73.3%) and NPV (82.9% vs. 71.1%).
Discussion
In this prospective study, we performed the first application of digital infrared thermal imaging to the detection of cervical lymph node metastasis from oral cavity cancer. We also found that the EGSVM-based infrared thermal imaging system is objective and reliable as a screening and diagnostic tool for the detection of cervical lymph node metastasis from oral cancer.
With manual qualitative analysis, we found that high frequency tumor-associated vascular abnormalities are powerful indicators of nodal involvement (Table 2). Compared with vessels in healthy tissues, tumor-associated vessels have long been observed to be abnormal in terms of morphology and structure, even at early stages of disease[]. In animal models, tumor-associated tortuous vessel morphologies appeared much earlier than a palpable mass, even when only tens of tumor cells were introduced into the tissue[]. As tumor growth needs an ever-increasing nutrient supply, malignant tumors generate a unique vessel system by creating new vessels (angiogenesis) and influencing major vessels around the tumor via angiogenic growth factors[]. Angiogenesis of the lymph node metastasis was imaged in a lymph node metastasis animal model over time[]. The tumor-associated vascular abnormalities, including increased vascular density with a tortuous vascular morphologic pattern, aberrant vasculature and a unilateral dilated vascular pattern in IR images, were consistent with previous studies. The metabolic activity and abnormal vessel pattern resulted in deviations in heat. It is this deviation in heat that provides the basis for the use of infrared imaging.
EGSVM is another major part of our work that contributes to the diagnostic performance. We have proposed an efficient computer-aided diagnosis system that uses the EGSVM model for classification. Computer-aided diagnostics have been studied in various diseases via medical images[]. A good computer-aided diagnosis system can eliminate operator dependency, improve diagnostic performance, and reduce the time needed for the interpretation of images. During manual qualitative analysis, abnormal signs are based on only relatively large areas that are visible to the naked eye, and thus, small lesions may be ignored. Automatic classification systems can describe the irregularity of pixels, allowing for greater objectivity in image processing and reduced inter-observer variability. Due to the satisfactory diagnostic performance, we anticipate that the proposed EGSVM will be a reliable and reproducible tool for the classification of thermal images.
The EGSVM-based infrared thermal imaging system has several advantages as a diagnostic tool for the detection of cervical lymph node metastasis from oral cancer. First, infrared thermal imaging is completely risk-free. It passively captures thermal radiation emitted by the human body, as there is no need for ionized radiation, contact, contrast agents or any invasive procedures. Second, compared with other expensive medical examinations, IR examination is low-cost due to the cheap IR camera that is used. In addition, IR image capture and automatic image analysis can be completed in less than one minute each, allowing results to be obtained quickly. Therefore, the EGSVM-based infrared thermal imaging system can also serve as a screening tool for cervical lymph node metastasis from oral cancer.
This study is a pilot clinical trial, and there are still limitations. We did not perform the comparison between CT, MRI and IR imaging. CT and MRI are commonly used for the detection of cervical lymph node metastasis. Further studies concerning this issue are necessary before IR imaging can be widely used for the detection of cervical lymph node metastasis from oral cavity cancer. Additionally, in this study, we extracted only one feature for automatic classification; the performance of the model can be improved with additional useful features.
The extending step is set (1/50 of width, 1/50 of height) in our feature extraction stage. Therefore, n is 50 in our experiments.