Oral cancer is a major health problem worldwide, causing over 127,000 deaths each year [1]. With an annual incidence exceeding 274,000 cases, oral cancer ranks as one of the top ten most common malignancies. Over two-thirds of cases and three-quarters of deaths due to oral cancer occur in developing countries [2]. In the U.S., the overall five-year-survival rate for patients with oral cancer is only 54%, one of the lowest rates of all major cancers; in developing countries, five-year survival rates drop below 30% [1, 3]. Patients with early lesions have better chances for cure and less treatment associated morbidity, yet despite the easy accessibility of the mouth, most patients present with advanced tumors, when treatment is more difficult, more expensive and less successful compared to earlier interventions. Early detection of oral premalignant lesions (OPLs) and early neoplastic changes may be our best and most cost-effective means to improve survival and quality of life for oral cancer patients from all socioeconomic communities [3].

No satisfactory mechanism exists currently to screen and detect early neoplastic changes of the oral cavity in the general population. Possible explanations include: 1) There is limited public awareness and insufficient education of health care workers about oral cancer risk factors, signs and symptoms. 2) Detection relies heavily on clinical experience at recognition of suspicious lesions during physical examination, with variable effectiveness [4, 5]. It can be difficult to distinguish OPLs from more common inflammatory conditions. 3) Practitioners and patients are reluctant to perform invasive biopsies of oral lesions which are expensive and often require referral to a specialist. 4) Some patients have field cancerization, where the entire mucosal lining that is exposed to carcinogens in alcohol and tobacco sustains damage and is at risk to develop cancer [6]. In high risk patients, often the whole lining of the oral cavity is potentially premalignant, making it difficult even for experienced clinicians to know when and where to biopsy [5]. The situation is even more challenging in developing countries and low-resource regions with high-risk populations, where a combination of lack of public awareness of the disease and inadequate resources and expertise for screening can result in even greater delays in diagnosis, leading to higher morbidity and mortality. One recent study conducted in southern India concluded that even simple visual examination by health workers can prevent 37,000 oral cancer deaths annually worldwide [7].

Recent advances in optical imaging have the potential to improve early detection of oral cancer and its precursors. Several groups have demonstrated that imaging systems that record the spatial distribution of tissue fluorescence at specific excitation/emission wavelength combinations can be used to survey large areas of oral cavity mucosa to non-invasively detect early changes associated with oral cancer in real-time [5, 7–12]. Lane presented a non-magnifying hand-held device for direct visualization of oral cavity tissue fluorescence, which is now FDA approved for clinical use and commercially available as the VELscope^® [5]. The system uses a metal-halide lamp with emission peaks at 405 and 436 nm to excite autofluorescence; images viewed by eye through the VELscope indicate a characteristic loss of fluorescence associated with malignant progression. Results from 50 biopsies taken from areas with loss of fluorescence in 44 patients showed a sensitivity of 98% and specificity of 100% for discriminating normal tissue from severe dysplasia, carcinoma in situ, or invasive carcinoma, using histology as the gold standard. An important finding was the ability of fluorescence visualization to aid clinicians in identifying early neoplastic lesions that were initially missed during traditional white light examination (WLE) [11].

To address the need for objective image interpretation, Roblyer recently described a multimodal digital microscope to obtain digital images of oral tissue fluorescence and reflectance and evaluated the ability of objective image analysis techniques to recognize OPLs and early oral cancer. Using 405 nm excited fluorescence images, algorithms could detect OPLs and oral cancer with a sensitivity of 94% and specificity of 87% in 64 patients [12]. The results of automated analysis of digital autofluorescence images must also be validated in larger clinical trials.

While digital image analysis can provide more objective interpretation, lower cost devices are needed for use in developing countries. To address the need for low-cost imaging, Rahman recently developed a multi-modal imaging system for screening and detection of oral cancer in high-risk populations in low-resource and remote settings [13]. This portable, battery powered system acquires near real-time digital images of oral tissue in reflectance and fluorescence mode. In this paper, we report results of a clinical study to evaluate the ability of this device to aid in detection of oral cancer in India involving 109 subjects at high-risk for developing oral cancer. We also characterize the optical properties of two types of potentially confounding oral lesions, melanosis and oral submucous fibrosis (OSF), which are common in this region, and discuss how the presence of these lesions influences the ability of the device to accurately identify neoplastic lesions.

Figure 1 shows typical white light reflectance and fluorescence images of sites corresponding to each of the four consensus clinical impression categories. Fluorescence images of sites with a consensus clinical impression of 'Normal' showed homogenous green mucosal fluorescence. Sites with a consensus clinical impression of 'Low Risk' or 'High Risk' exhibited a progressive loss of green fluorescence, while sites with a consensus clinical impression of 'Cancer' typically showed both a loss of green fluorescence and an increase in orange-red fluorescence.

Figure 2 shows white light reflectance and fluorescence images of two potentially confounding lesions found in the study population - oral submucous fibrosis (OSF) and melanosis. In white light reflectance images, OSF sites exhibited a pale white, patchy appearance similar to leukoplakia. However, unlike leukoplakia which is often associated with decreased green fluorescence, no loss of fluorescence was observed in sites with a clinical description of OSF. Melanosis sites were easily recognized from the white light images as areas of dark pigmentation and were consistently associated with decreased green fluorescence.

Figure 3 shows scatter plots of the image features and corresponding ROC curves for classification algorithms which resulted in the best sensitivity and specificity, respectively. Figure 3a shows the normalized MFI from the blue channel by diagnostic category for all sites. Those sites with a consensus clinical impression of 'Normal' and clinically visible melanosis or OSF are shown separately. Sites with a consensus clinical impression of 'Low Risk', 'High Risk' or 'Cancer' exhibit decreased normalized blue MFI compared to 'Normal' sites. Sites with melanosis show decreased normalized blue MFI, while those with OSF have similar normalized blue MFI to other 'Normal' sites. Figure 3b shows the ROC curve for the algorithm based on normalized blue MFI, excluding melanosis from the analysis. The operating point indicated on the ROC curve corresponds to a sensitivity of 92% and a specificity of 84%; the threshold values corresponding to this operating point is a normalized blue MFI of 0.86, which is indicated by the solid horizontal line in Figure 3a.

Figure 3c shows a scatter plot of the normalized ratio of red to green MFI by consensus clinical impression for all sites. Sites with a consensus clinical impression of 'Low Risk', 'High Risk' or 'Cancer' exhibit an increased normalized red to green MFI ratio compared to 'Normal' sites; on average, the normalized red to green MFI ratio increases as the severity of the consensus clinical impression increases from 'Low Risk' to 'Cancer'. Sites with melanosis show an increased normalized red to green MFI ratio, while those with OSF have a similar normalized red to green MFI ratio as other 'Normal' sites. Figure 3d shows the ROC curve for the classification algorithm based on the normalized red to green MFI ratio, excluding melanosis from the analysis. The operating point indicated on the ROC curve corresponds to a sensitivity of 90% and a specificity of 87% relative to the gold standard of consensus clinical impression. The threshold value corresponding to this operating point is a normalized red to green MFI ratio of 1.11, and is indicated by the solid horizontal line shown in Figure 3c.

This pilot study demonstrates that objective analysis of fluorescence images obtained with a low-cost imaging system can classify sites as neoplastic or non-neoplastic with high sensitivity and specificity relative to the gold standard of consensus clinical impression. The performance of this low-cost, objective system compares favorably to results reported for pilot studies of other optical imaging systems. A recent review by De Veld summarizes several clinical studies of optical imaging with qualitative image analysis for detection of oral neoplasia and reports sensitivity ranging from 63% to 100% and specificity from 79% to 96% [14]. More recently, Lane et al. achieved a sensitivity of 98% and specificity of 100% using qualitative assessment of fluorescence images acquired with the VELScope to discriminate dysplasia and cancer from normal oral mucosa [5]. Roblyer reported quantitative analysis of fluorescence and reflectance images obtained with a multi-spectral digital microscope to classify oral premalignant lesions and oral cancer with a sensitivity of 94% and specificity of 87% [12]. Our results, obtained in a South Asian population using the low-cost, portable optical imaging system described here, demonstrate similar sensitivity and specificity.

It should be noted that most studies of optical imaging based adjunctive techniques for oral cancer detection have been carried out in high prevalence populations, and recent reviews stress the need for randomized controlled studies in low prevalence populations [15, 16]. To be useful in low resource settings where the vast majority of oral cancers occur, screening aids must be affordable and must not rely on significant clinical expertise for image interpretation. Our results demonstrate the feasibility of conducting these much needed trials using low-cost, portable imaging devices in low resource settings.

The optical characteristics of oral lesions found in this South Asian population are similar to those in other studies. Our finding of decreased fluorescence associated with neoplastic lesions is consistent with reports in the literature. This loss of autofluorescence has been attributed to a decrease in collagen cross-links associated with neoplastic transformation [17]. Our results also indicate a relative increase in red fluorescence for neoplastic sites. Other studies have made similar observations, attributing this increased red fluorescence to porphyrins [8, 9, 17].

In addition, we found the normalized red to green MFI ratio to be amongst the top three best-performing features. Roblyer also reported that this parameter was the best performing feature for quantitative identification of oral premalignant lesions and oral cancer using a multispectral digital microscope [12]. Figure 4 presents the normalized red to green MFI ratio for 57 sites in 21 patients measured by Roblyer at 450-nm excitation; results are similar to those measured in this study (Figure 3c). Although the two studies were conducted independently on two entirely different populations and using two different instruments, the threshold ratio values from the two different systems, nevertheless, were very similar. The threshold to differentiate neoplastic sites from non-neoplastic sites is in good agreement between the two studies; here we found a threshold of 1.11 was optimal to differentiate neoplastic sites from non-neoplastic sites and Roblyer found a threshold of 1.09 could differentiate premalignant lesions and cancers from normal tissue.

The ability of fluorescence-based screening aids to differentiate precancerous lesions from benign lesions such as inflammation must also be validated in larger trials and the optical properties of other potentially confounding lesions, which may be population specific, must be characterized. For example, betel quid use is common in south Asia and is associated with a high incidence of OPLs as well as other potentially confounding oral lesions, including melanosis and oral submucous fibrosis (OSF) [18, 19]. A recent study of 130 patients in the United States noted that 72% of lesions clinically characterized by inflammation or pigmentation show loss of fluoresecence with VELscope examination [20].

We characterized the optical properties of two types of potentially confounding lesions, melanosis and OSF, specific to the geographic region. Melanosis is usually benign and not considered to be precancerous [19]. Our results indicate that sites with melanosis exhibit decreased fluorescence, but can easily be recognized by their characteristic appearance in white light reflectance images. The loss of autofluorescence in melanosis is likely due to strong absorption of light by black pigmentation in the superficial epithelium. In contrast to melanosis, the malignant transformation rate of OSF has been estimated to be between 3% and 19% [21]. Most of the OSF sites measured in this study were graded as 'Low Risk'. We found that the OSF sites imaged here did not exhibit loss of fluorescence. Histologically, OSF is characterized by juxtaepithelial fibrosis, along with atrophy or hyperplasia of the overlying epithelium, keratinizing metaplasia, and accumulation of hyalinized collagen beneath the basement membrane [22–24]. All of these constituents are strong sources of autofluorescence and may contribute to autofluorescence observed in our images of OSF sites. The role of combined reflectance and fluorescence imaging with quantitative image analysis for better discrimination of inflammation and neoplasia should be further explored.

While the sensitivity and specificity reported here are encouraging, there are a number of limitations of this study. First, the same dataset was used both to develop classification algorithms and to assess their performance; in this situation, potential over-training can inflate estimates of sensitivity and specificity. Results must be verified in an independent validation set. Second, the gold standard used to assess algorithm performance was consensus clinical impression; due to resource limitations, histopathologic diagnosis was not available from all sites. Finally, a large number of 'Low Risk' sites were misclassified as neoplastic by the optical algorithms presented here. It is interesting to note that 'Low Risk' sites with consensus among all three expert observers were more likely to be classified by the optical algorithms as non-neoplastic (6/10 = 60%) than were 'Low Risk' sites where only two of the expert observers agreed (13/30 = 43%). Additional studies with histologic endpoints for all sites are required to assess the ability of quantitative optical image analysis to aid in the evaluation of low risk oral lesions.

The clinical study presented here demonstrates the ability to identify neoplastic and non-neoplastic oral tissue in vivo objectively using the low-cost, portable imaging system. Although further work is needed to address the limitations in this study, results from this pilot study suggest that this simple imaging device can potentially improve oral screening efforts in low-resource settings where clinical expertise and resources are often limited.