Radiomics model of 18F-FDG PET/CT imaging for predicting disease-free survival of early-stage uterine cervical squamous cancer

2.1Patient

This study was approved by the ethics committee of Fudan University Shanghai Cancer Center. We retrospectively enrolled 201 uterine cervical squamous cancer patients who underwent PET/CT scans 1 week prior to radical surgery between January 2010 and June 2015. Patients were excluded if they received neoadjuvant chemo- or radiotherapy or had other malignancies or other histologic types of uterine cervical cancer. All patients were diagnosed with IB-IIA stage (FIGO 2009). Postoperative adjuvant radiochemotherapy was determined by pathologic risk factors according to NCCN and FIGO guidelines [4, 5].

Postoperative pathologic features included tumor size, histologic type, stromal invasion depth, LVSI, parametria involvement, positive surgery margin, and lymph node metastasis in our study. According to guidelines [4, 5], parametrial involvement, positive surgery margin, or pelvic lymph node metastasis mentioned by postoperative pathology were high-risk factors. Cervical tumor diameter ⩾ 4 cm, stromal invasion depth ⩾ 1/2, or lymphovascular space invasion (LVSI) were intermediate-risk factors. Patients with at least one high-risk factor were defined as patients with high-risk disease. Patients with at least one factor of intermediate risk were defined as having intermediate-risk disease. Patients with none of these risks were defined as having low-risk disease. All pathological slides were independently reviewed by two experienced gynecologic pathologists according to the WHO criteria [17]. Immunohistochemical staining was performed for Ki67 (Roche 30-9) using a Ventana Benchmark XT autostainer (Ventana Medical Systems Inc., Tucson, AZ, USA).

The surveillance is based on the patient’s risk for recurrence and personal preferences by the recommendation of NCCN guidelines: clinicoradiological and physical examinations were routinely recommended every 3–6 months for the first 2 years, every 6–12 months for another 3 to 5 years, and then annually. Patients with pathologic high-risk disease can be assessed more frequently (for example, every 3 months for the first 2 years) than patients with low-risk disease (e.g., every 6 months). All patients were followed up until June 30, 2019. Disease-free survival (DFS) was defined as the time interval from the date of surgery to the date of recurrence or metastasis.

Figure 1.

Outline of the workflow from data input, feature extraction, selection, and model construction.

Squamous cell carcinoma antigen (SCCA) is a very specific tumor marker that exists in the cytoplasm of cancer, such as uterine, cervix, lung, head and neck squamous cell carcinoma. SCCA is not only an early indicator of response during cervical squamous cancer therapy but also a prognostic biomarker after therapy. High pretreatment SCCA is associated with recurrence [18, 19, 20]. Ki67 is a nuclear protein, and its expression level is strongly associated with tumor cell proliferation and growth and is widely used in routine pathological investigations as a proliferation marker. Clinically, Ki67 expression is significantly higher in malignant tissues with poorly differentiated tumor cells than in normal tissue [21, 22]. In cervical cancer, the expression of Ki67 is closely related to the occurrence and development of cervical carcinoma [22]. Therefore, the expression of SCCA and Ki67 levels were evaluated in our study for prognostic research. Finally, 16 clinicopathological data types, including age, preoperative SCCA level, FIGO stage, Ki67 level, above pathologic features, recurrence/metastasis status, DFS, and conventional PET/CT parameters, were collected from medical records, cancer registries, and pathology reports. In total, 201 patients were included in the study and were randomly divided into a training cohort (n= 140) and a testing cohort (n= 61). The clinicopathologic characteristics of all patients are shown in Table 1.

2.218F-FDG PET/CT image acquisition and interpretation

2-Deoxy-2-(18F)fluoro-D-glucose (18F-FDG) was produced automatically by cyclotron (Siemens CTI RDS Eclipse ST, Knoxville, Tennessee, USA) using the Explora FDG4 module in our center, and the radiochemical purity was over 95%. All patients fasted for at least 4–6 h before 7.4 MBq/kg (0.2 mCi/kg) 18F-FDG injection and were quiet for approximately 1 h before scanning. Venous blood glucose levels were maintained under 10 mmol/L in our center. Images were obtained on a Siemens biograph 16HR PET/CT scanner (Knoxville, Tennessee, USA). The transaxial intrinsic spatial resolution was 4.1 mm (full-width at half-maximum) in the center of the view. CT scanning (120 kV, 80–250 mA, pitch 3.6, rotation time 0.5) from the proximal thighs to the head was first performed for data acquisition, followed by a PET emission scan. The acquisition time was 2–3 min/bed. PET image data sets were iteratively reconstructed using the attenuation correction of CT data, and the infused images were displayed on a workstation.

All PET/CT images were evaluated by two experienced nuclear medicine physicians independently. Consensus was reached in case of discrepancy. A multimodality computer platform (Syngo; Siemens) was used to analyze the 18F-FDG PET/CT images. The reconstructed images were then converted to a semiquantitative image corrected by the injection dose and the participant’s body weight (SUV), which was calculated as [decay-corrected activity (kBq) per milliliter of tissue volume]/[injected 18F-FDG activity (kBq) per gram of body mass]. SUVmax was calculated by placing a spheroid-shaped volume of interest (ROI) within the primary cervical tumor. MTV and SUVmean were measured by drawing contours large enough to encase the tumor in the axial, coronal, and sagittal PET images. A common threshold value of SUV of 2.5 was used to define the margins around the target. The voxels presenting an SUV intensity of greater than 2.5 within the contouring margin were incorporated to define the MTV, and the mean voxels within this contouring were measured as the SUVmean. Ureters and bladder were manually subtracted to ensure correct voxels. The TLG was calculated according to the following formula: TLG = SUVmean × MTV. Conventional PET/CT parameters, such as SUVmax, SUVmean, MTV, and TLG, were not obtained from radiomic feature extraction, so we incorporated these indicators into clinical characteristics.

2.3PET/CT image segmentation and feature extraction

The workflow is illustrated in Fig. 1. The ROI of images was delineated independently by two experienced nuclear medicine physicians using ITK-SNAP software (Version 3.6, United States). After manual segmentation, the PyRadiomics package [23] was adopted to automatically calculate the radiomic features from tumor ROIs, including shape, intensity and texture [24]. Shape features could reflect the geometric properties of tumors and were only calculated on original images. Intensity features, which describe the voxel intensity distribution, were calculated using first-order statistics. Texture features were calculated based on different texture matrixes to quantify tumor heterogeneity.

For CT images, in addition to the original images, a wavelet filter and a Laplacian of Gaussian (LoG) filter were applied to obtain derived images for richer feature extraction [25]. The wavelet features were extracted based on wavelet decomposition and reflected multiscale information of intensity and texture. LoG filters were used to enhance edge information and emphasize tumor textures of different coarseness levels. In total, 1218 radiomic features were extracted from CT images.

For PET images, feature calculation was only performed based on original images considering the low-resolution characteristics of PET images, and the personal dose and weight were used for normalization. A total of 100 PET features were extracted.

Figure 2.

Selected radiomic features in predicting DFS. Feature importance of radiomic (A) and clinicopathologic features (B). f1, CT_wavelet-HLH_gldm_SmallDependenceHighGrayLevelEmphasis; f2, CT_log-sigma-4-0-mm-3D_glrlm_RunVariance; f3, CT_log-sigma-4-0-mm-3D_gldm_LowGrayLevelEmphasis; f4, CT_log-sigma-5-0-mm-3D_glrlm_LongRunHighGrayLevelEmphasis; f5, CT_log-sigma-5-0-mm-3D_gldm_DependenceNonUniformity; f6, PET_original_glcm_SumSquares; f7, TLG; f8, LVSI; f9, lymph node metastasis; f10, deep stromal invasion; f11, preoperative SCCA level; f12, FIGO stage.

Finally, we obtained a total of 1318 radiomic features from CT and PET images.

2.4Feature selection

Radiomic feature selection was performed in the training set in the following three steps. First, the concordance index (C-index) was used to evaluate the univariate prognostic ability of each feature, and a low C-index indicated poor predictive ability for prognosis. Features with a low C-index were removed. Second, the variance inflation factor (VIF) was applied to quantify the collinearity between features and remove redundancy from high-dimensional features [26]. In each iteration, the feature with the highest VIF value was removed. This procedure was performed iteratively until the VIF values of all remaining features were below a certain threshold. Third, the random survival forest (RSF) [27] model was used for multivariate analysis. RSF is an ensemble of tree-based learners that can evaluate the importance of input features. For each feature, an importance score was calculated by RSF based on evaluations of relevance and prognosis between all features. Then, features with important scores below a preset threshold were removed, and the radiomic signature was generated using the remaining features. For a total of 16 clinicopathologic features, considering the small number of features, only the last two steps were performed. The model parameters and thresholds were determined by grid search, and all procedures were implemented using stats-models [28], scikit-learn [29] and scikit-survival packages [30, 31, 32].

2.5Prognostic model establishment

Based on the selected features, the popular ensemble RSF method was used to build the prognostic model. The radiomic signature and clinicopathologic features were first enrolled in the RSF model independently and then fused into a comprehensive model. For all of the above models, 4-fold cross-validation was performed in the training set to determine the model parameters.

2.6Statistical analysis

Continuous variables were described as medians with ranges, and categorical variables were described as frequencies with percentages. Training and testing set characteristics were compared using the descriptive statistical analysis method for continuous variables and χ2 or Fisher’s exact tests for categorical data. Kaplan-Meier analysis was performed to evaluate the predictive capacity of the model. The performance of the prognostic model was evaluated on the independent testing set with the C-index, which is most frequently used in survival analysis and can assess the overall prognostic ability of the model. To further evaluate the performance of the model in a specific time range, the time-dependent cumulative/dynamic AUC (C/D AUC) was calculated, which is an extension of the area under the receiver operating characteristic (ROC) curve in survival analysis with censored data [33, 34, 35]. The mean value of the C/D AUC (m C/D AUC) was used as an indicator to measure the performance of the preoperative predictive model. All statistical tests were two-sided, and p< 0.05 was considered statistically significant.

3.1Clinicopathologic characteristics

As shown in Table 1, the clinicopathologic characteristics of the training and testing cohorts were similar. A total of 201 patients were divided into two groups: a training cohort with a median age (range) of 47 (19–73) years and a testing cohort with a median age (range) of 48 (25–74) years. The FIGO stage ranged from IB1-IIA2.

In the training cohort, the median (range) DFS was 56 (1–111) months. A total of 19 (13.6%) patients relapsed or metastasized during the follow-up period. The median (range) preoperative SCCA was 2.2 (0.3–70.0) ng/ml. Over half of the patients had postoperative pathologic risks, including 62 (44.3%) with tumor size ⩾ 4 cm, 102 (72.9%) with stromal invasion depth ⩾ 1/2, 58 (41.4%) with LVSI, 6 (4.3%) with parametrial invasion, 2 (1.4%) with positive surgery margins, and 46 (32.9%) with lymph node metastasis. Eighty-six (61.4%) patients underwent postoperative adjuvant therapy. The immunohistochemical Ki67 value was recorded with a median of 75% (range 20%–95%). For PET/CT scans, the median (range) routine PET/CT clinical parameters were as follows: SUVmax, 12.58 (2.65–37.41); SUVmean, 5.17 (2.57–12.77); MTV, 26.43 (0.41–235.56); and TLG, 143.70 (1.05–1806.75).

Figure 3.

The K-M curve of radiomic, clinicopathologic, and combined model in training and testing dataset.

In the testing cohort of 61 patients, 8 (13.1%) patients relapsed or metastasized with a median DFS of 52 months (range 3–113 months). The median (range) SCCA level and Ki67 value were 3.4 (0.3–70.0) ng/ml and 80% (20%–95%), respectively. The tumor size was larger than 4 cm in 38 (62.3%) patients. In total, 45 (73.8%) patients were pathologically proven to have a stromal invasion depth ⩾ 1/2, 23 (37.7%) to have LVSI, 4 (6.6%) to have parametrial invasion, 1 (1.6%) to have a positive surgery margin, and 18 (29.5%) to have lymph node metastasis. For PET/CT parameters, the median (range) SUVmax, SUVmean, MTV and TLG were 12.82 (2.50–71.84), 5.38 (2.72–20.56), 37.00 (0.28–263.37) and 230.27 (0.76–2183.34), respectively.

3.2Feature selection

The important radiomic features were selected from a total of 1318 radiomic features, including 1218 CT features and 100 PET features. First, univariate analysis was performed, and the results showed that 178 CT features and 22 PET features were considered to be predictive and were selected. Then, the collinearities of the remaining features were analyzed. Twenty-five independent features (21 CT features and 4 PET features) were selected when the redundant features with VIF values larger than 10 were removed. Finally, with the multivariate analysis based on RSF, 6 important radiomic features were selected, including 5 CT features and 1 PET feature. As shown in Fig. 2A, CT-wavelet-HLH_gldm_ SmallDependence HighGrayLevelEmphasis and CT-log-sigma-4-0-mm-3D_glrlm_RunVariance were the most important features compared with others. The definitions of 6 selected radiomic features are presented in Table 2, which were in compliance with the Imaging Biomarker Standardization Initiative (IBSI) [36].

For 16 clinicopathologic features, 5 redundant features were first removed by VIF. Then, the features with a nonpositive important score in RSF-based multivariate analysis were further removed, resulting in 6 prognostic clinicopathologic features (Fig. 2B), including TLG, LVSI, lymph node metastasis, deep stromal invasion, preoperative SCCA level, and FIGO stage.

3.3Performance of the prognostic model

After feature selection, the RSF model was established to predict the risk scores for each patient. For each model, patients in the training and testing sets were stratified into high- and low-risk groups by the optimal cutoff value obtained from the log-rank test on the training set. The cutoff values of the radiomic, clinicopathologic and combined models were 4.9, 3.9 and 4.0, respectively. The Kaplan-Meier curves of DFS revealed a significant difference between the high- and low-risk groups (Fig. 3).

The overall performances of the models were evaluated with the C-index value. As illustrated in Table 3, the PET/CT radiomic model yielded robust discrimination, with C-index values of 0.9557 for the training set and 0.9338 for the testing set, which was better than the clinicopathologic model (0.9125 for the training set and 0.9019 for the testing set). By merging the radiomic and clinicopathologic features, a comprehensive model was further built and achieved C-index values of 0.9717 and 0.9527 for the training and testing sets, respectively.

Figure 4.

The C/D AUC curve of the radiomic, clinicopathologic, and combined models in the testing dataset. As an indicator to measure the preoperative predictive efficiency of the model in addition to the C-index, the mC/D AUC was further calculated over all time ranges. The curve showed that the combined model achieved the best performance during the research period, and the mC/D AUCs of the radiomic, clinicopathologic and combined models were 0.9146, 0.8645, and 0.9199, respectively. At 60 months after surgery, the preoperative combined model showed a higher predictive value with a C/D AUC60 of 0.9703 compared to 0.9248 in the clinicopathologic model and 0.8923 in the radiomic model (tag line shown). However, the predictive ability was obviously decreased after 100 months.

The cumulative/dynamic AUC (C/D AUC) curve was used to evaluate the model performance in a specific time range. The curve showed that the combined model achieved the most robust performance during the research period (Fig. 4). The mC/D AUC, which is likely to be the C-index, was calculated to measure the predictive efficiency of the model over all time ranges in Table 3. Compared to single PET/CT radiomic (mC/D AUC = 0.9146) and clinicopathologic models (mC/D AUC = 0.8645), the combined model outperformed a better prognostic prediction (mC/D AUC = 0.9199) in the testing set. Furthermore, a higher C/D AUC value at a specific time t indicates that the model could better distinguish patients who suffer relapse/metastasis before time t from those after time t. We found that the preoperative combined model showed a high predictive value, with a C/D AUC value (C/D AUC60) of 0.9703 at the time point of 60 months after surgery, compared to 0.9248 in the clinicopathologic model and 0.8923 in the radiomic model (shown in Table 3 and Fig. 4). The predictive ability was obviously decreased after 100 months.