Evaluation of prognostic models developed using standardised image features from different PET automated segmentation methods

This is a retrospective cohort study of consecutive patients with biopsy-proven OC, including gastro-oesophageal junctional (GOJ) tumours, radiologically staged between 16 September 2010 and 31 July 2016. Patients were identified from a database of OC patients used in a previous study [15]. Institutional Review Board approval was granted and requirement for informed consent was waived (Wales REC 1, UK reference 14/WA/1208).

Overall, 486 patients with FDG-avid primary oesophageal and GOJ tumours were considered for inclusion. Fourteen patients were excluded due to missing clinical data. All patients were deemed to have potentially curable disease following contrast-enhanced CT staging investigation. All PET/CT examinations were performed separately, following the initial CT, and reported in the same centre by Consultant Radiologists with an interest in Nuclear Medicine. Radiological staging was performed according to the International Union Against Cancer (UICC) TNM 7th edition [16]. Following exclusions, 472 patients were studied.

Patients were fasted for at least 6 h prior to tracer administration. Serum glucose levels were routinely checked and confirmed as less than 7.0 mmol/L prior to imaging. Patients received a dose of 4 MBq of ¹⁸F-FDG/kg. Uptake time was 90 min, standard practice at our institution. A GE 690 scanner (GE Healthcare, Buckinghamshire, UK) was used. CT images were acquired in a helical acquisition with a pitch of 0.98 and tube rotation speed of 0.5 s. Tube output was 120 kVp with output modulation between 20 and 200 mA. Matrix size for the CT acquisition was 512 × 512 pixels with a 50 cm field of view. No oral or intravenous contrast was administered. PET images were acquired at 3 min per field of view. The length of the axial field of view was 15.7 cm (skull base to mid-thigh). Images were reconstructed with the ordered subset expectation maximisation algorithm, with 24 subsets and 2 iterations. Matrix size was 256 × 256 pixels, using the VUE Point™ time of flight algorithm. All PET-based data was obtained using the same PET/CT scanner and reconstruction method with voxel dimensions of 2.73 × 2.73 × 3.27 mm.

Patients began treatment 2–4 weeks after staging FDG PET/CT imaging. Patients either had endoscopic mucosal resection (EMR), surgery alone, neo-adjuvant chemotherapy (NACT) or neo-adjuvant chemoradiotherapy (NACRT) prior to surgery, definitive chemo-radiotherapy (dCRT) or palliative therapy. The optimum treatment strategy was decided by consensus at the MDT. In general, fit patients with tumours pre-operatively staged as T3/T4a, N0/N1 were pre-operatively treated with NACT or NACRT. Less fit patients, or those with T1/2 N0 disease, had surgery alone. Patients deemed unsuitable for surgery due to co-morbidity and/or performance status, extensive loco-regional disease, or personal choice received dCRT.

Manual delineation of the metabolic tumour volume (MTV) is limited by intra- and inter-observer variability and is time consuming [17‐19]. Semi-automated and automated segmentation methods are favourable alternatives by reducing variability in delineation and decreasing the contouring time [20]. Fixed percentage thresholding has been shown to be dependent upon the SUV_max of a tumour as well as the MTV [21]. Furthermore, it has been shown that texture analysis of PET imaging is dependent upon the segmentation method used to define the MTV [12, 22, 23]. However, more complex segmentation algorithms such as adaptive iterative thresholding (AT) have been shown to be independent of SUV_max as well as being correlated to the MTV. Segmentation methods adopting clustering techniques such as Fuzzy C-means (FCM), Gaussian fuzzy C-means (GCM) and K-means (KM) using 2, 3 and 4 clusters (FCM2, GCM3–4, KM2 - KM4), as well as region growing (RG) and watershed transform (WT) methods, are promising segmentation methods in the delineation of the MTV. These segmentation methods are reviewed in detail in the report by Hatt et al. [9], are described in detail previously [24] and are summarised in Table 1. In each case, the MTV was defined using AT, FCM2, GCM3, GCM4, KM2, KM3, KM4, RG and WT PET segmentation methods.

A clinical radiologist subjectively assessed each tumour contour produced by all nine PET segmentation methods for accurate tumour representation. All tumour contours were visualised using the same software and image settings to ensure consistent methodology. Segmentation methods were considered inadequate for further analysis if less than 90% of contours were non-representative. This pre-defined value was decided upon prior to image visualisation. Contours were assessed individually and classified as not representative if contours were greatly different from the primary tumour, or included bone, lung or medistinial tissue. In addition, segmentation methods that had failed and conformed to the boundary of the bounding box were defined as not representative of the primary tumour.

Only primary tumours were analysed to ensure consistent methodology across all patients. Before quantitative image analysis and texture feature extraction, PET images were re-sampled into 0.5 SUV bins. A fixed bin width maintains a constant intensity resolution when compared to approaches based on a fixed number of bins [25]. In the development of the prognostic models, age at diagnosis (number of years), radiological stage (stage IA-IV) and treatment (curative vs palliative) were included because these are strong predictors of survival [26]. Curative and palliative treatments were coded as 1 and 2 respectively. Radiological staging was modelled categorically.

Radiomic analysis was performed using features implemented as part of the Image Biomarker Standardisation Initiative (IBSI), a multicentre, international collaboration aimed at improving the reproducibility and validation of quantitative medical image analysis studies [5]. The radiomic features selected for inclusion in this study were chosen as they have shown prognostic and predictive significance in other radiomic studies investigating OC [12, 27, 28]. These have been summarised in Table 2. Moreover, many radiomic feature implementations have been described [6, 7, 27, 29] and are divided into three groups for which a summary is provided. In this study, the MTV was analysed as a 3D volume with no thresholding applied to the MTV mask.

First-order statistical metrics summarise the voxel intensity distribution within the segmented MTV, without concern for spatial relationships [30]. First-order metrics are typically histogram based and reduce the MTV to singular values describing the mean, minimum, maximum, median, and uniformity of the intensities within the MTV. Included in first-order stastical analysis is Skewness (asymmetry measure), Kurtosis (pointiness measure) and Entropy (randomness measure). Kurtosis and skewness have been shown to be independent predictors of survival [15] and of prognostic significance in the literature [31].

Higher-order statistical metrics retain spatial information and are used to quantify inter-voxel intensity relationships. Dissimilarity is the quantification of variation in voxel pairs and is calculated using a Grey Level Co-occurrence Matrix (GLCM) generated for each unique direction and averaged. A low dissimilarity is resultant of neighbouring voxels having similar values [32]. Zone percentage is calculated from a Grey Level Size Zone matrix (GLSZM) by assessing the fraction of recorded zones compared to the maximum number of possible zones. Heterogeneous MTVs have high zone percentage scores. Grey Level Non-Uniformity (GLNU) is an evaluation of the distribution of zone counts for each intensity value. The feature value is low when the number of zones associated with each intensity value are similar. Coarseness is a neighbourhood grey-tone difference matrix (NGTDM) feature that gives an indication of the level of spatial rate of change in intensity [33]. GLCM, GLSZM and NGTDM can be computed in 2D or 3D. The matrices in this study were computed in 3D as this may highlight the multi-scale, directional properties of tumour tissue [34].

The primary outcome of the study was OS, defined as number of months survived from date of diagnosis. Patients were followed up 3-monthly for the first year, 6-monthly until 5 years then annually thereafter, or until death. All included patients were followed up for at least 12 months. Date of death was obtained from the Cancer Network Information System Cymru database (CaNISC, Velindre NHS Trust, Wales).

The final steps of each prognostic model are presented in Table 4. Three known clinical prognostic factors (age, radiological stage and treatment) remained in each derived model, but there was a difference in the inclusion of texture metrics by segmentation technique. AT and KM2 produced the same model output. Interestingly, IBSI metrics were not included in the final models for these segmentation methods. However, skewness and kurtosis were independently significant for survival using GCM3 method. Skewness and GLNU were significant using WT method. Their inclusion in the models illustrates their additional prognostic value compared with current prognostic factors.

The equations for each model derived from different segmentation methods were used to calculate the prognostic scores, and are listed in Table 5. These calculations were derived using published methods [35].

Figure 1 shows the risk stratification for WT, KM2, AT and GCM3. Median OS for the low-risk, intermediate-risk and high-risk groups in the AT- and KM2-derived prognostic model was 36.0 months (29.9–42.1 months), 18.0 months (15.1–20.9 months) and 9.0 months (7.8–10.2 months), respectively. Median OS for the low-risk, intermediate-risk and high-risk groups in the GCM3-derived prognostic model was 36.0 months (28.8–43.2 months), 18.0 months (15.4–20.6 months) and 9.0 months (7.7–19.2 months). Median OS for the WT derived prognostic model low-risk, intermediate-risk and high-risk groups was 36 months (27.8–44.2 months), 19 months (15.1–23 months) and OS for the high-risk group was 9 months (7.7–10.3 months) respectively. Table 6 shows the number of patients stratified as low, intermediate and high-risk for each single prognostic model along with the prognostic score range for each risk stratification group. Table 7 shows the number of patients whom change risk stratification.

The largest proportion of patients to change risk stratification group was between prognostic models based on GCM3 and on WT (n = 73, 17.1%). It can be noted that no patient changed risk stratification group between AT and KM2 because the prognostic models were identical. The number of concordant patients stratified as low, intermediate and high-risk across the developed models was 118 (28%), 95 (22%) and 116 (27%), respectively. There was no overall survival difference between AT, GCM3, KM2 or WT low-risk groups (χ² 0.052, df 3, p = 0.997), intermediate-risk groups (χ² 0.016, df 3, p = 0.999) or high-risk groups (χ² 0.028, df 3, p = 0.999).

For interest, Additional file 1 describes the developed prognostic models for the excluded PET-AS methods. Additional file 2 describes variances in radiomic features extracted using differing discretisation methodologies, which is an important consideration in radiomic studies. Additional file 3 describes the correlation of MTV with the extracted radiomic features.