Discovery of pre-therapy 2-deoxy-2-¹⁸F-fluoro-D-glucose positron emission tomography-based radiomics classifiers of survival outcome in non-small-cell lung cancer patients

Lung malignancy is a leading cause of cancer-related death, with a predicted 5-year survival rate of 8–13% [1]. Worldwide, approximately 1.8 million new cases were diagnosed in 2012. Distinct from histology, stage, and performance status, the ability to provide prognosis on the basis of tumor biology is often lacking in current clinical practice. More recently, DNA sequencing from several tumor regions has been undertaken to highlight spatio-temporal mutational heterogeneity [2, 3]. Currently, imaging identifies the sites of disease and response to treatment by assessing the change in size but other than TNM staging provides limited prognostic information. In addition, outcomes of patients within each TNM staging group can vary widely highlighting the need for more accurate prognostic markers. Potential interventional methods to assess genetic heterogeneity will probably employ multi-core invasive biopsy, which limits its safe use for routine prognosis determination. The micro- and macro-structure of tumors, however, also harbor heterogeneous phenotypes due to factors such as hypoxia, necrosis, directional/non-directional tumor cell growth, vascular density, and immune infiltration. It is hypothesized that the asymmetric local, regional, and global density and architectural distortions of tumor phenotypes could have prognostic value, and this has resulted in a new 'omics' — radiomics [4‐6] — whereby quantitative features describing tumor phenotypes are extracted in high-throughput from routine radiologic images and further processed by machine learning methods for prognostication; such high-dimensional output of tumor phenotypic heterogeneity is thought to have important prognostic value, with drug resistance and potential for development of metastatic spread implied.

2-deoxy-2-¹⁸Fluorine-fluoro-D-glucose positron emission tomography-computed tomography (FDG-PET/CT) is routinely used for staging lung cancer prior to consideration of radical treatment such as surgery or chemo-radiotherapy including the use of stereotactic body radiotherapy. Indeed, radiomics classifiers based on the CT component have been investigated for predicting lung cancer histology and shown to have moderate prediction accuracy [7]. Beyond the use of FDG-PET/CT for staging, we investigated in the present study whether pre-therapy radiomics features derived from routine FDG-PET/CT examinations of non-small-cell lung cancer (NSCLC) patients who were subsequently treated with radiotherapy/chemo-radiotherapy across multiple hospitals might harbor useful prognostic information.

This multi-institution retrospective study showed that a radiomics feature vector, FVX, derived from analysis of FDG-PET data of primary tumors ≥ 5 ml in patients with NSCLC is invariant to PET scanner properties and predicts OS. Accurate prognostic information is crucial in stratifying newly diagnosed NSCLC patients to different treatments or best supportive care. Currently, TNM staging is the primary method to stratify treatment approach in NSCLC; however, it offers imprecise prognostic information, leading to both under-treatment and over-treatment in some patients. Other established prognostic factors for lung cancer include performance status (Karnofsky or ECOG (Eastern Cooperative Oncology Group) classification), weight loss (e.g., > 5%) and systemic inflammation (C-reactive protein or modified Glasgow Prognostic Score) [18‐20]. While factors such as EGFR (epidermal growth factor receptor) mutation predict response to targeted therapy [21‐23], tumor-specific prognostic factors are lacking. In the current work, we assessed the role of radiomics features as prognostic factors in NSCLC. A machine-learning-enabled weighted linear combination of the statistical features of GLSZM_SzVarianc_64gl and NGTDM_Complex_64gl — FVX — was found to possess prognostic information and importantly was invariant to scanner properties investigated (Supplementary Table 7). The features do not have immediate physiological relevance. GLSZM is a regional ‘homogeneity’ texture feature that calculates lengths of uniform pixels (picture elements) in a 2D image, or in our case, directionally-independent groups of uniform voxels (volume elements) in each of the 26 available directions in 3D space; GLSZM_SzVarianc_64gl (size variance of the GLSZM at 64 Gy level) examines the variance in the number blocks by size (independent of the gray-level) and is negatively correlated with survival, possibly identifying hypoxic or necrotic regions with poor prognosis [24‐26]. NGTDM represents contrast, and is determined by examining changes in intensity between a target voxel and the surrounding neighbors to enable calculation of apparent difference between neighboring regions of voxel intensities. Contrast is related to the information content of an image and is a mathematical measure of heterogeneity; non-responding tumors with poor prognosis tend to have higher contrast [27]. NGTDM_Complex_64gl (complexity of the NGTDM for 64gl), which refers to the average visual complexity within the volume, is positively correlated with survival, although with less of a magnitude than SzVarianc, and perhaps acting as a balance on SzVarianc.

This is one of the first reports of a whole tumor image-derived lung cancer prognostic factor. In our analysis, there was a higher hazard of death (1.74; p = 0.00116) when the median FVX was used as input. The implicit assumption here is that a set of mathematically-derived tumor phenotypes correlate with survival. It should be noted, however, that death as an endpoint could have occurred by other means — indirect consequence of the tumor or non-tumor related — or may have been subjected to different variations of physician choice of chemotherapy/chemo-radiotherapy, making this analysis the more interesting. Furthermore, routine PET variables, while useful for staging, did not possess prognostic information. Baseline primary tumor SUV_max has been reported by some groups, but not all, to predict outcome in NSCLC patients. For resectable NSCLC patients, a meta-analysis of 13 studies showed that primary tumor SUV_max has significant prognostic value on patient survival [28]. A more recent meta-analysis, assessing the prognostic value of primary tumor SUV_max prior to radiotherapy in NSCLC, reported that higher tumor SUV_max was correlated with shorter OS, particularly in stage I NSCLC receiving stereotactic body radiotherapy (SBRT) [29].

Volumetric parameters, such as MTV and TLG, which consider the whole tumor volume, have been reported to be prognostic in NSCLC. Secondary analysis of the large multicentre prospective American trial of 196 inoperable Stage IIb/III NSCLC has reported MTV, TLG to be strongly prognostic for OS, while SUV_max was not [30]. The TLG and MTV, which were shown in our study to highly correlate with each other, were not significant on multivariable analysis.

As this was a large multi-centre study acquired on different scanners, voxel sizes differed (Supplementary Table 2); we did not standardise the voxel sizes. However, slice thickness and matrix size did not significantly affect the FVX. The robustness of FVX, invariant to instrument factors including slice thickness, permits this variable to be applied in multi-institutional studies. Previous work on scanner types [31‐33] have yielded mixed results in terms of texture stability across model and manufacturer type, although limited models have been used. In addition, much thought has been given in the methodology to reduce Type 1 errors and false discovery which have entered the published literature [34]. Compared to CT technology that has seen substantive reduction in slice thickness, PET FWHM (full width at half-maximum; a measure of resolution) has not seen such substantive change over the past decade, and this could have led to the scanner invariance of our study. We set a threshold of 5 ml, in keeping with earlier work of Soussan [8]. It is likely that inclusion of smaller tumors would have led to higher variability given the poor resolution of PET (compared to CT or MRI). Hence, the inferences from this study are limited to the group of patients presenting with medium-large lesions. The classification of patient subgroups in the Kaplan–Meier analysis was based on the FVX value calculated from combination of weighted radiomic features. FVX was a continuous variable, and we have demonstrated that FVX was linearly correlated with overall survival in the training and validation sets using a continuous Cox regression analysis. It should be noted that other analyses were performed: Cox regression based on dichotomised median FVX, Kmeans clustering and optimal cut-off (Youden’s J) from the ROC curve, and showed consistent results (data not shown).

TESTI was an unusual dataset in terms of size and heterogeneity with a high number of stage 1 tumors and high mortality. The fact that these patients had radiotherapy ± chemotherapy instead of surgery indicates that there was probably associated poor performance status. Unfortunately, information on performance status was not available. However it was felt important to attempt to test the radiomics signature against this unusual dataset, as ideally the ‘real-life’ prediction of the radiomics feature vector should work irrespective of sample size.

A recent single-institution study of a PET/CT radiomics signature for prediction of disease-free survival (DFS) in NSCLC undergoing surgery with curative intent reported that image derived parameters outperformed TNM staging in predicting DFS. However, although promising, this was a single-centre study, utilising what appears to be unenhanced CT scans without external validation and in a different cohort of patients to our study [35].

Limitations of the present study should be highlighted. 1) This is a retrospective, albeit multi-institutional, study and future prospective studies in similarly large cohorts will be needed to verify this novel endpoint. 2) We did not consider other prognostic factors as these were not consistently available from all institutions. Addition of other prognostic factors in future studies will enable more rigorous assessment of events likely to have caused death. 3) We did not consider EGFR mutations/expression or other genetic outcomes, as these were not available for all patients. 4) Approximately 14% of the initial 535 patients screened were excluded for having tumours < 5 ml (Supplementary Table 1). The choice of 5 ml reflects a statistical limitation of applying radiomics to PET data (less data-points within the VOI compared to CT). Thus, with regard to generalisation of our study to patients having chemo-radiotherapy, we would caution the exploitation of our findings to smaller tumors. Irrespective of these limitations, we highlight a massive opportunity for physicians and patients, whereby mathematically-derived features from scans that newly diagnosed NSCLC patients would normally have as part of routine care can be ‘re-purposed’ to predict prognosis. Only software implementation and computing power are required for incorporation into patient management pathways; thus, we envisage easy acceptance of this potentially cost-effective methodology for use with existing prognostic methods. As it is tumor-specific, patients stratified to poor prognostic FVX groups could be candidates for earlier follow-up or a lower threshold in change of therapy [36].

In summary, we have discovered a scanner-invariant radiomics feature vector that performs well in independent validation and test datasets. This multi-institutional study provides new opportunities for prospective assessment of radiomics features for prognosis in patients with NSCLC.