In conventional screening methods, pathologists typically identify early cancer by assessing morphological markers, such as variation in nuclear size and shape, staining intensity, chromatin pattern, and the appearance of nucleoli or vacuoles and mitotic figures. These methods are highly time-consuming and subjective; furthermore, accuracy is heavily dependent on the expertise of the pathologist 
. In this study, we present a novel approach for automated detection of early cancer in tissue based on simultaneous quantification of mitochondrial organization, cellular morphology, metabolic activity, and keratin localization from non-invasively acquired TPEF images. This method is capable of assessing morphological changes that are typically evaluated by pathologists, but also allows for quantification of mitochondrial organization and metabolic activity of tissue, which are both potentially valuable diagnostic indicators of disease not available to pathologists using traditional screening methods.
To develop and assess the validity of the analysis methods described in this study, we exploited tissue engineering approaches that allowed us to develop organotypic epithelial tissue cultures that mimic in many respects the biochemistry, architecture and organization of normal and pre-cancerous squamous epithelia 
. We used the HKc/DR cell line as a model for HPV-associated high-grade premalignant lesions. This clonal line of HPV16 immortalized human foreskin keratinocytes was selected for resistance to differentiation and failure to growth arrest in response to TGF-ß 
. HKc/DR cells are non-tumorigenic but show patterns of gene expression changes that are similar to cervical carcinoma cells 
. Consistent with earlier studies that showed that HPV immortalized keratinocytes exhibit histological abnormalities similar to high-grade premalignant lesions 
, organotypic cultures of HKc/DR cells grew in a disorganized fashion and exhibited a low level of differentiation throughout the various depths similar to what was observed in the histological images (). In contrast, organotypic cultures of primary HFKs exhibited layers of distinct morphology and a keratinized uppermost layer, similar to what is typically seen in healthy tissues. Since the sample is optically sectioned, the tissue remains intact after optical biopsy, and therefore many areas can be interrogated in vivo without inducing pain. The 3-D optical biopsies can be easily viewed from any angle/position in space and quantitatively analyzed using PSD approaches to assess the morphology and mitochondrial organization and multi-wavelength image analysis to assess metabolic activity and keratin localization.
Morphological assessment using PSD analysis is faster and includes more cells than conventional segmentation techniques that have been developed for analysis of stained histological sections 
, and, more recently, for images acquired using non-invasive techniques, such as confocal microscopy 
. The goal of many of these studies was to develop methods to rapidly process images and obtain quantitative cellular measures, such as cell border irregularity and nuclear to cytoplasmic ratio, as a means to discriminate between normal and diseased tissue. Although several of these algorithms proved sufficient to identify cells grown in culture, there remains a significant challenge to accurately identify the often ill-defined cell borders in tissue. The result of poor border definition is often the exclusion of many cells from the analysis. To overcome the challenges associated with accurate segmentation, we adopted a Fourier-based approach to assess morphology over many length scales. These methods are sensitive to the spatial TPEF intensity variation patterns from the sub-cellular to the multi-cellular level. Since we are not thresholding or segmenting the images, all cells are included in the analysis.
Our analysis reveals that the lack of variation in nuclear size as a function of depth, which is one of the histopathological hallmarks of intraepithelial neoplasia, is a parameter that can be extracted by examining the variance of the PSD at spatial frequencies that correspond to nuclear sizes. Based on size estimations of the dark appearing areas within the cytoplasm, which should correspond to the non-mitochondria containing nuclei, we have found that the nuclei in the tissues we examine are approximately 8–10 µm in diameter ( insets) and, therefore correspond to the spatial frequencies around 0.1 µm−1
. As shown in , this is a spatial frequency regime of high variance for the normal tissue equivalents, but not for the HPV tissues. In addition, the PSD can be used to assess more subtle and largely unexploited tissue features, such as the level of mitochondrial organization 
. Assuming self-affine fractal organization, we can extract the Hurst parameter as a quantitative measure of mitochondrial organization to differentiate between normal and precancerous tissue (p<0.05).
There have been a number of studies that use fractal-based characterization for pathological assessment of tissue from the cervix 
, stomach 
, mammary glands 
, breasts 
, and oral cavity 
, while other diagnostic applications of fractal analysis have been reviewed elsewhere 
. In our work, we calculate the 2-dimensional PSD from label-free images that are not binarized. This is particularly adventageous when alnalyzing tissue images, where the source of contrast is inherently dim endogenous signals, and difficulties arrise when trying to spatially segment (binarize) images based on intensity. In many other studies that use fractal based analyses, excised tissue is stained to enhance the contrast before imaging 
. With the high contrast agent the images are easily binarized to enhance the cell borders, thus creating a 1-D, self-similar fractal pattern from which the fractal dimension is computed, typically using box counting methods. The 1-D fractal dimension gives insight into the irregularity of cell borders and it has been used to exploit differences between nuclear borders of normal and cervical intraepithelial neoplastic cells 
. There have also been studies that assess the fractal dimension in conjunction with conventional morphological measurements and report an increased sensitivity over morphological assessment alone 
The values of the Hurst parameter used in our analysis can vary between 0 and 1, corresponding to the highest levels of anti-correlation and correlation, respectively, with a Hurst value of 0.5 corresponding to Brownian fractal behavior. The Hurst parameter values that represent the mitochondrial organization of the epithelial cells we examine are consistently lower than 0.5, with the lowest values corresponding to more highly differentiated cells (). Using similar 2-D, self-affine fractal analysis, Schmitt and Kumar also report anti-correlated fractal patterns from phase contrast images of liver tissue 
. Also consistent with our findings, Einstein et. al. observe increased levels of anti-correlation of the fractal patterns of chromatin density in breast cell nuclei associated with cancer 
In addition to the rich morphological and organization information that can be extracted from analysis of the intrinsic NAD(P)H fluorescence intensity fluctuations, TPEF images acquired at a combination of excitation and emission wavelength bands can be analyzed to acquire quantitative information about metabolic activity and keratin content and localization. To establish the number of components and corresponding spectral emission profiles of the dominant contributors to the detected TPEF images from our tissues, we initially analyzed a series of spectrally resolved images acquired from 400 to 700 nm at 755, 800 and 860 nm excitation. This analysis revealed significant contributions from three chromophores, whose emission spectra were consistent with NAD(P)H, FAD and keratin fluorescence 
. Unfortunately, depending on the excitation and emission wavelengths where TPEF is detected, there can be significant spectral overlap between the chromophores, which complicates the extraction of quantitative biochemical information. Guided by analysis of the full spectral emission profiles, we determined that it was important to segment the TPEF images in keratin-positive and keratin-negative regions. Since the fluorescence intensity from keratin is on the order of 3 to 6 times greater than NAD(P)H and FAD 
, we achieved this by a simple thresholding procedure based on the sum of the blue and green detection channel intensities at 755 nm and 860 nm excitation, respectively. We then focused on the keratin(−) pixels and found that in the suprabasal and basal layers we could isolate the contributions of NAD(P)H and FAD from analysis of the images detected by the 455 nm channel at 755 nm excitation and the 525 nm channel at 860 nm excitation with 85–90% confidence, as assessed by comparison with the full spectral decomposition results.
Reliable extraction of the NAD(P)H and FAD fluorescence was important for characterizing the metabolic profile of the normal and dysplastic tissues. In particular, we used the redox ratio, whose value has been shown to be inversely proportional to metabolic activity 
. An advantage of the redox ratio assessment of tissue compared to individual estimates of NADH or FAD is that artifacts potentially introduced by absorption of chromophores, such as keratin, are minimized, since they similarly affect NADH and FAD. We find that the redox ratio decreases as we move from the superficial/cornified to the suprabasal cell layers of the normal epithelial tissues and then it increases slightly as we reach the basal layer. In comparison, we observed significantly less change in the redox ratio depth profile of the HPV-immortalized tissues (p<0.05), consistent with the apparent loss of differentiation. Such differences in the depth-dependent variation of metabolic activity between normal and pre-cancerous tissues have been reported by other investigators 
. Furthermore, we report that the overall metabolic activity is higher in the pre-cancerous tissues than in normal tissues (p<0.05), consistent with previous studies 
. More recently, researchers observed that the redox ratio variations are also correlated with the degree of metastatic potential and location within the tumor 
While elimination of the fluorescence contribution from keratin was necessary to accurately assess metabolic activity, we also used the keratin localization as a means to discriminate between tissue types (p<0.05). In the normal tissues, we found that keratin, likely type 10 and 13 from terminally differentiated cells 
, was restricted to the upper layers and consistently expressed throughout the depths of the HPV tissues. Indeed, keratin expression has been correlated with cervical cancer progression 
, with keratin 17 prominently expressed at various tissue depths of dysplastic tissue 
. Due to the high quantum yield of keratin 
, it could serve as an additional diagnostic biomarker than can be much more easily extracted from analysis of intrinsic TPEF compared to conventional histological images, where identifying keratin amongst cells is not trivial 
The biochemical, morphological and organizational information extracted from analysis of the intrinsic TPEF images is complementary and can be used in combination to ultimately develop highly sensitive and specific algorithms for the detection of pre-cancerous lesions. Using multivariate linear regression, we found that such a combination provided a better means to identify precancerous changes in tissue when analyzed simultaneously. We specifically evaluated and considered the Hurst range (p<0.05), redox ratio (p<0.05), keratin localization (p<0.05), and variance of the PSD (p<0.10). Addition of the keratin localization did not improve significantly the diagnostic potential over the combined use of the other three parameters in the small sample size of this study. It is important to note, however, that these studies are to be regarded as a proof of principle, where we harnessed organotypic tissue culture of a clonal HPV16 immortalized cell line that was previously established as a model system for HPV-associated high-grade premalignant anogenital tract lesions. To establish the true diagnostic potential of the parameters that we derived from this study, it will be important to validate these on human tissue samples, where a much more heterogeneous cell population will be expected. Most premalignant lesions contain areas with different levels of cellular abnormalities. Low-grade lesions, in particular, often retain at least some ability to undergo differentiation. Improvements in diagnostic accuracy using a combination of morphological and biochemical information extracted typically through the use of multiple optical modalities has been observed previously, consistent with our findings 
. A unique aspect of the approach presented in this study is that all the information is extracted from analysis of a single type of imaging with approaches that can be implemented in real time. In addition, the procedures include all cells in a given field and require no human input, unlike commonly used segmentation approaches. Since the methods are based on non-invasive measurements, the algorithms could be useful in a screening/diagnostic device that can assess appreciable tissue areas and detect the early microscopic changes associated with pre-cancer development relying on information that clinicians do not currently have access to.