Fractal analysis of nuclear histology integrates tumor and stromal features into a single prognostic factor of the oral cancer microenvironment

Almost 30 000 individuals are diagnosed with oral squamous cell carcinoma (OSCC) each year in North America and approximately 6000 of these patients succumb to the disease, annually [1]. OSCC is an aggressive disease and even favourable treatment outcomes are associated with significant morbidity. Five-year survival rates for OSCC have remained between 40 and 50% for the past several decades. Biomarkers that can identify aggressive disease at diagnosis and inform treatment decisions might improve survival outcomes and quality of life for OSCC patients. Although several prognostic markers for OSCC have been described in the literature, treatment is directed predominantly by the tumor-node-metastasis (TNM) staging system.

The tumor microenviroment is a dynamically interacting entity composed of tumor cells and the surrounding stroma. These interactions are not only critical for tumor growth and progression but also for treatment sensitivity/resistance. Therefore, effective prognostic biomarkers should ideally incorporate features of both tumor and stroma, leading to a more comprehensive assessment of tumor biology. Histopathological features have been previously used for prognostication in OSCC. Brandwein-Gensler and colleagues described a histologic risk assessment score based on pattern of invasion (POI), perineural invasion (PNI) and lymphocytic infiltration (LI) [2,3]. These authors propose that achieving negative resection margins do not guarantee local disease-free and overall survival benefits. On the other hand, a combination of histopathological features of the tumor (POI and PNI) and stroma (LI) accurately predicted risk of local recurrence and survival. Although the manual assessment of histological features is a powerful technique for predicting risk, it requires expert subject knowledge of head and neck histopathology and can be very labour intensive. Since few diagnostic laboratories have access to specialized head and neck pathologists, integration of histological feature analysis into a single, digital histopathology-based biomarker may improve the utility and encourage clinical adoption of this type of prognostic testing.

Fractal dimension (FD) is a mathematical measure of the irregularity and complexity of a shape and may be used for the digital assessment and quantification of histological features in the tumor microenvironment [4]. In contrast to our intuitive notion of dimension (i.e. the topological dimension), which is an integer value (0 for a point, 1 for a line, 2 for a plane, etc.), the FD can be a non-integer value that is greater than the topological dimension. The extent to which the FD of an object may be greater than the topological dimension depends on the space filling capacity of the object. FD is a non-integer number that quantifies the degree of space filling of an object. True mathematical fractals exhibit a higher degree of space filling because they exhibit exact or statistical self-similarity in structural patterns when examined to infinitely small scales. As such, actual fractals do not exist in nature, since there is a fundamental natural limitation to the scaling behaviour of natural objects [5]. However, FD analysis has found widespread use in medical image analysis because it lends itself naturally to the pragmatic characterization of irregular non-Euclidean structures found in medical images [6,7]. One such application of FD has been to discriminate the architectural complexity of biological structures associated with neoplastic states. Previous studies have applied FD analysis for the diagnosis, staging and prognosis of several cancer-types including breast [4,8], prostate [9,10], colon [11], lung [12], endometrial [13], gall bladder [14], larynx [15] and OSCC [16]. FD analysis of nuclear histology digitally quantifies the space filling properties of nuclei. Such analysis, when performed on the entire tumor microenvironment (tumor and stroma) can be a source of rich prognostic information.

We have previously used fluorescence immunohistochemistry (IHC) and automated quantitative analysis (AQUA®) to investigate the prognostic value of proteins associated with apoptosis [17,18], proliferation [19] and hypoxia [20,21] in OSCC. We have also reported that the prognostic impact of these biomarkers differs according to their cellular distribution within the tumor microenvironment. Tissue microarrays (TMAs) used to examine protein biomarkers are ideally-suited for digital histological analysis since TMA cores contain both tumor and stromal tissue compartments. When performing AQUA®, images of each whole TMA core are generated and the nuclei are routinely co-stained with 4′,6-diamidino-2-phenylindole (DAPI) to differentiate nuclear/cytoplasmic localization of a biomarker. In this study, we computed the fractal properties of DAPI-stained nuclei in whole TMA cores. We hypothesized that nFD analysis will integrate tumor and stromal characteristics commonly incorporated in OSCC histopathological risk assessment methods into a single, prognostic factor for OSCC. We report that nFD is a robust and powerful independent prognosticator of patient outcome that integrates the proliferative properties of the tumor compartment and the immunologic properties of the stromal compartment into a unified prognostic entity that is amenable to clinical translation.

We report a digital histopathologal image-based prognostic biomarker (nFD) derived from fractal analysis of DAPI-stained nuclei. This single measure integrates features of both the stromal and tumor compartments in the tumor microenvironment. nFD can effectively discriminate between OSCC patients with good and worse prognosis and was an independent prognostic indicator in our OSCC cohort when the model was adjusted for established prognostic clinical covariates. A strong positive correlation was observed between nFD and a pathologist-scored assessment of LI in the stroma. nFD was also positively correlated with proliferation (scored using fluorescence IHC and AQUA®), an important tumor-associated prognostic marker. The significance of both these features, independently, suggests that nFD scores are an effective method for the automated, image-based, integration of both stromal and tumor features with acknowledged prognostic value.

OSCC is a serious public health problem worldwide and the lack of effective prognostic biomarkers adversely affects patient management and survival outcomes. This has led researchers to look beyond the traditional TNM staging system and investigate biological correlates to histopathological features of the tumor microenvironment. Staining with DAPI is a routine component of IHC that helps identify nuclei. We hypothesized that FD analysis of DAPI-stained images (computer-acquired when performing AQUAnalysis®), could provide a digital histology-based prognostic factor for OSCC that might be more objective and less labour-intensive than traditional histopathological analysis. FD analysis has been previously used in OSCC. Several researchers have demonstrated that FD can discriminate between normal versus malignant oral tissue [27,28]. Goutzanis and colleagues have used FD to assess vascularization and also nFD as a prognostic factor [16,29]. However, contrary to our results, these authors report that high nFD is associated with poor prognosis [16]. It is worth noting that almost all previously reported studies have used 3,3′-dichlorobenzidine (DAB) IHC-based images for FD analysis. Also, these studies did not take into account the tumor microenvironment that might provide valuable biologic information relevant to prognosis. DAPI-based nuclear staining is more robust than DAB IHC-based techniques since it is not affected by antibody specificity issues. Also DAPI staining is relatively easy to perform that protein-based IHC since protein is more sensitive than DNA to pre-analytical variables, particularly when the motive is to preserve overall DNA structure rather than specific base pairs. Also, DAPI staining allows for multiplexing of diverse stains, allowing for staining of additional proteins that can, e.g. discriminate the tumor (PCK) and stromal (vimentin) compartments [21]. Interestingly, we found that nFD scores from the stromal or tumor compartment alone did not show a significant association with survival. However, a robust association with survival was observed when nFD from the whole tissue core (tumor plus stroma) was considered.

In order to understand how well digital nFD-based histopathological analysis correlates with expert pathological assessment of stromal morphological parameters associated with survival, we compared LI, scored by a pathologist, with nFD scores. We also studied the relationship between nFD and tumor proliferation in order to evaluate if nFD correlated with tumor-associated prognostic features. LI has been previously reported to be associated with prognosis in OSCC [24-26] and was significantly associated with both DSS and OS in our OSCC cohort as well (Additional file 1: Figure S1). We observed that high nFD was associated with increased LI in the stroma and cell proliferation in the tumor. We believe that high nFD tumor and stroma might reflect increased proliferation in the tumor and the presence of infiltrating immune cells in the stroma. Both have been shown to be associated with improved prognosis attributed to increased susceptibility to radiotherapy. We found that patients with high nFD scores have significantly better survival when treated with post-operative radiotherapy compared to patients with low nFD (Figure 3b). Although further investigation is required, nFD analysis may represent a summary measure incorporating tumor proliferation and the immune involvement that predicts response to radiotherapy [30]. Therefore, high nFD might represent a state where high proliferation renders tumor cells sensitive to radiotherapy and dying cells are cleared by the infiltrating immune cells, creating a “perfect storm” for the tumor.

The histopathological scoring system proposed by Brandwein-Gensler et al. is a powerful tool but its broad implementation may be limited by its labour intensiveness and the requirement for extensive training. Additionally, consistency of scoring is a substantial risk given the high degree of inter-observer variability in other oral histopathology-based systems, among pathologists [31,32]. Digital pathology-based nFD scoring incorporates multiple biomarkers and therefore might provide a more reliable and objective indicator of prognosis compared to single biomarker-based assays. We believe that a comprehensive approach to the analysis of tumor microenvironment, such as the one presented here, will improve prognostication and outcomes in OSCC.