1 Introduction
As the differential diagnosis of pulmonary nodules and mass lesions in many countries, the positron emission tomography (PET) with 18F-fluorodeoxyglucose (18F- FDG) has been established. According to a meta- analysis of the published data on 18F-FDG–PET scanning from January 1996 to September 2000, the average sensitivity and specificity of 18F-FDG–PET scanning for detecting a malignancy was 97% and 78%[1]. The 18F-FDG–PET has not the benefit of evaluating pulmonary nodules with low 18F-FDG avid[2]. Compared with 18F-FDG–PET, the combination of 18F-FDG–PET with the diagnostic CT scan without intravenous contrast can improve the sensitivity from 69% to 97%, and does not change the specificity[3]. The diagnostic accuracy of 18F-FDG PET/CT without the quality CT of lung nodule was similar with that of 18F-FDG–PET[4,5]. Nie et al.[6] reported a semiautomatic computer-assisted diagnostic (CAD) scheme, indicating that the combination of 18F- FDG–PET with CT can differentiate benign from malignant pulmonary nodules, and is better than 18F- FDG–PET or CT alone. In that study, we excluded two nodules with extrathoracic malignancy, and focused on lung lesions in the whole body 18F-FDG–PET. Lesion outside of lung cancer avid for 18F-FDG is considered as metastasis. The extrapulmonary lesion is avid for 18F-FDG, indicating that the newly diagnosed lung lesion is malignant.
2 Methods and materials
2.1 Patient cohort
With a lung lesion newly diagnosed at the conventional chest radiography or the X-ray CT for a suspicious malignant primary, the patients underwent whole body integrated 18F-FDG PET/CT from February 2005 to January 2009. We selected the pathological confirmation and suspicious infection patients cured by antibiotic, and excluded the patients with recent malignancy history and diabetes mellitus.
We gained the approval of Institutional Review Boards without patient informed consent.
The 199 consecutive patients (63 women and 136 men, 25–88 years old, mean age±SD: 58.8±12.4) met the further analysis criteria. In 205 lung lesion patients, their biopsy or surgery proved that the 169 was malignant; and 32, benign in pathology; and 4 infection cured by antibiotics (Table 1).
| Malignants | Numbers | Benigns | Numbers |
| Adenocarcinoma | 101 | Tuberculosis | 11 |
| Squamous cell carcinoma | 34 | Sarcoidosis | 2 |
| Adenosquamous carcinoma | 4 | Organized pneumonia | 4 |
| Large cell carcinoma | 1 | Abscess | 4 |
| Small cell carcinoma | 10 | Cryptococcal pneumonitis | 1 |
| Synoviosarcoma | 1 | Infection/inflammation | 10 |
| Hodgkin's lymphoma | 1 | Other benign lesions were not further classified | 4 |
| Metastatic carcinoma from rectum/ lung | 3 | ||
| Unspecified NSCLC | 14 | / | / |
| Total number | 169 | / | 36 |
2.2 Acquisition of PET/CT image
To fulfill the blood glucose level of less than the 8.3 mmol/dL, patients were asked to fast over 6 h and rest for 15 min before administration 5-MBq/kg 18F-FDG. The images were acquired by a whole-body PET/CT scanner (Discovery DST 8; GE Healthcare, USA) after injection at 60 min. The whole body scan was approximately from the middle thighs to the skull roof. The CT scan without intravenous contrast was used as a protocol involving 140 kV, 150 mA, 0.8 s per tube-rotation, and 3.75-mm slice thickness. The PET scan was performed with a 3.27-mm section thickness, 3.5 min per table position, and two-dimensional acquisitions. Patients were asked to maintain normal shallow respiration during image acquisition.
The 18F-FDG-PET images were reconstructed using CT attenuation correction and an ordered subset expectation maximization algorithm. The maximum standardized uptake value (SUVmax) was determined by Ref.[7].
2.3 Acquisition of CT images
All patients underwent a breath-hold spiral CT scan on the lung lesions without intravenous contrast by 120kV, 170 mA, 0.8 second per tube-rotation, 2.5-mm slice thickness, and 1.35-pitch after PET/CT scan.
2.4 Interpretation of CT images
The CT images were interpreted by two radiologists, who were unaware of each patient's history and PET images. Each lesion was described by its site, size, attenuation, shape, margin characteristics, consolidation, cavitation, and invasion. The criteria for interpreting lesions were applied[5]. Readings in case of differing results were performed in consensus.
2.5 Interpretation of PET/CT images
The 18F-FDG–PET criteria for malignancy were used as citation[8]. The locations of abnormal tracer uptake were recorded, and the metastasis index (MI) was scored by a 5-points scale (Table 2). The PET/CT images were analyzed by two doctors with experience of 5 years in the 18F-FDG PET/CT. Also readings were conducted in consensus in case of differing results.
2.6 Statistical analysis
The patient characteristics were compared by the Student’s t-test and the chi-squared test. When the expected values in the any cells of the contingency table were below five, the Fisher exact test was conducted. Variables reaching the significance might be included in a multivariate logistic regression model. Selection method with entry testing was based on the significance of the score statistic; and removal testing, the probability (P>0.15) of the likelihood ratio test[9]. Odds ratios (OR) and the 95% CIs were computed by unconditional logistic regression. A receiver-operating -characteristic (ROC) analysis included the CT interpretations, and the lesion uptakes in SUVmax, and the predicted malignancy probability (P) which was calculated by the fitted multivariate logistic regression model. The difference in the area under curve (AUC) was tested by the Z statistic. A 2-tailed p values (≤ 0.05) showed statistical significant differences. The diagnostic OR for a test is defined as sensitivity/(1‒ sensitivity) ×specificity/(1‒specificity)[10].
3 Results
3.1 Results of CT and PET/CT
The patient characteristics and their lesions are shown in Table 2. In the 58 lesions, their lesion sizes could hardly be determined because of the local atelectasis and its associated obstruction, and parahilar location. In the 147 lesions, the lesion sizes were 36±21 mm in the range of 4–143 mm). Locations of distant lesions avid for 18F-FDG are listed in Table 3.
The non-contrast CT shows that 62% (104/169) of malignant lesions were classified as positive; and 29% (59/205) of all lesions, equivocal; and 44% (16/36) of benign lesions, negative (Table 2). Taking equivocal lesion as malignancy, the diagnostic OR was 8.2; and as benignancy, 3.6.
In Table 2, the sensitivity for 18F-FDG PET/CT was 95% (161/169) and the specificity was 25% (9/36) at the diagnostic OR of 6.7. The sex, age, and lesion sizes were no obvious difference between benign and malignant lesions, but the CTscore, SUVmax, 18F- FDGscore, and MI were highly different (p<0.01).
| Characteristic | Malignancy | Benignancy | p | |
| Patients | Women/Men | 50/113 | 13/23 | 0.58 |
| Age / year | 59±13 | 57±12 | 0.29 | |
| Lesions | Size (mm)a | 37±22(n=122) | 31±20 (n=25) | 0.21 |
| CTscore (0/1/2) | 15/50/104 | 16/9/11 | <0.01 | |
| 18F-FDGscore(0/1) | 8/161 | 9/27 | <0.01 | |
| SUVmax | 10.7±6.0 | 5.7±3.7 | <0.01 | |
| MI(0/1/2/3/4)b | 48/11/34/15/61 | 21/1/7/3/4 | <0.01 |
| Location/number | Malignancy | Benignancy |
| Celiac lymph nodes | 8 | 2 |
| Retroperitoneal lymph node | 8 | 1 |
| Pelvic lymph nodes | 2 | 1 |
| Inguinal lymph nodes | 1 | 1 |
| Lymph node of neck | 12 | / |
| Axillary nodes | 4 | / |
| Tonsil | 1 | / |
| Skeleton | 26 | 1 |
| Liver | 5 | 1 |
| Suprarenal gland | 7 | / |
| Brain | 3 | / |
| Cervical cord | 1 | / |
| Thyroid gland | / | 1 |
| Abdominal wall | 1 | / |
| Subcutaneous soft tissue | 2 | / |
| Gluteus | 1 | / |
| Pulmonary trunk and left atrium | 1 | / |
| Total number | 83 | 7 |
3.2 Multivariate stepwise logistic regression model analysis
By stepwise selection, CTscore, SUVmax, and MI were included in the fitted multivariate logistic regression model without 18F-FDGscore. In this model, when increasing, either CTscores or SUVmax or MI was used as malignancy predictors. The P was calculated by P=ex/ (1+ex), where x = –1.16+0.87(CTscore)+0.15 (SUVmax) + 0.27(MI) (Table 4).
| Variables | Coefficient (β±SD) | Odds Ratio (95% confidence interval) | p |
| CTscore (0/1/2) | 0.87±0.27 | 2.39(1.41–4.05) | <0.01 |
| SUVmax | 0.15±0.05 | 1.16(1.04–1.29) | <0.01 |
| MI(0/1/2/3/4) | 0.27±0.14 | 1.31(1.00–1.73) | 0.05 |
| Constant | –1.16±0.47 | 0.31(0.12–0.79) | 0.01 |
The model estimated that the P was more than 50% at x≥0. The CTscore calculated by the equation was 2 at x>0. In other words, at the positive CT scan, the lung lesion should be taken as malignancy regardless of 18F-FDG uptake[3]. For example, an adenocarcinoma in the right upper lobe was noted, its CTscore was 2; and SUVmax, 1.5; and MI, 0; and P, 69.1%, as shown in Fig.1.
At the same time, when CTscore was 1 and SUVmax was not less than 2, the x was larger than zero. To put it at the indeterminate CT scan, the SUVmax =2.0 was used as cutoff for malignancy[3,11], thus helping to distinguish between benign and malignant lesions. When CT score was zero, the x was larger than zero at the SUVmax of larger than 7.7, suggesting that the lung lesion should be considered as malignancy regardless of its morphological information. When diagnostic results of CTscore were contradictive with the SUVmax, the MI could help for differential diagnosis. For example, an inflammation lesion in the left lower lobe was observed, its CTscore was zero; and SUVmax, 5.0; and MI, zero; and P, 39.9% (Fig.2).
3.3 ROC analysis
A ROC analysis for CTscore, SUVmax and P was performed (Fig.3). The SUVmax curve based on the lung lesions has the cutoff of initial 1.5, step 0.5, and end 8.0. The SUVmax curve intersects the CTscore curve. The additional values on CTscore and SUVmax curves hold over the entire range of sensitivity and specificity. The AUC was 0.83±0.04 for the model; and 0.71±0.05, for CTscore; and 0.74±0.05, for SUVmax. Statistical analysis by the AUC shows that the model was superior and significantly different from SUVmax (p=0.01) and CTscore (p<0.01). There was no significant difference between SUVmax and CTscore at p=0.64. At the cutoff value of 0.50 for malignancy, the sensitivity was 96% (163/169); and specificity, 31% (11/36); and the diagnostic OR for P, 12.0.
4 Discussion
In the 18F-FDG–PET, the SUVmax at OR=1.16 (p<0.01) was associated with lung malignancy in the model (Table 4). The SUVmax increased as a significant malignancy predictor was similar to Grgic’s report [10]. The sensitivity of 18F-FDG PET/CT scanning for detecting malignancy was 95%, this was equivalent to the available data of 97% sensitivity by meta- analysis[1], and a prospective multicenter study of 92% sensitivity[8]. But the 25% specificity in our study was lower than those of the reported 44%–85%[1,3,4,5,8,11,12,13]. Also, the benign lesions, including tuberculosis, bacterial pneumonia, organized pneumonia, active sarcoidosis, infectious granulomas, and acute pyogenic abscesses, and so forth, can produce false-positive readings in 18F-FDG–PET[14], these were mainly caused by the low specificity.
The CT morphological information was an effective predictor for malignancy at OR=2.41(p<0.01), and its sensitivity and specificity were not as accurate in this study as in a multicenter of contrast-enhanced CT (98% sensitivity, and 58% specificity)[15]. As we know, a lung lesion at CT diagnosis suspicious for a malignant primary is usually dependent on morphological information and enhancement[16]. Some equivocal lesions may shift to definitely malignant or definitely benign on its enhancement, and improve the CT sensitivity and specificity. The less accuracy was likely caused without contrast-enhanced CT[15].
The MI was included in the model at OR=1.31(p=0.05). To the best of our knowledge, the MI is the first proof whether lesions outside of lungs avid for 18F-FDG might help to distinguish the malignant lung lesion. Based on the score statistic probability (p=0.05), we should be cautious to make a conclusion by MI.
The 18F-FDGscore was excluded from the model because its information had been included in its SUVmax. The stepwise logistic regression was designed to find the most parsimonious set of predictors in predicting malignancy.
The ROCs show that the model lies over the curve for CT diagnosis and SUVmax, indicating that the model accuracy is superior regardless of where setting the threshold defining a malignancy test. Furthermore, the statistical test by the AUC shows that the superior model was different from CTscore and SUVmax. Compared with the diagnostic OR of 3.7 or 8.3 in different diagnosis protocol for CT and 18F-FDG (6.3), the diagnostic OR reached the highest value of 12.0 at the cutoff of 0.50 P, suggesting the model was better than CT and SUVmax alone. The model had potential to develop a semiautomatic CAD to aid and improve physician diagnostic skills, and it would be tested by a multicenter in the future.
5 Limitation
In this retrospective clinical study, our patients were selected from their larger pool, and referred to the 18F-FDG PET/CT study because of the pathologic verification of their lung lesions, these induced a selection bias. The malignancy rate was 82% (169/205). No matter how there were several influencing factors such as serum glucose level, respiration, partial-volume effects, and noises[17], the SUVmax was in use as the de facto standard[18].
6 Conclusion
In this study, the extrapulmonary lesions avid 18F-FDG could distinguish the malignancy from benignancy. Our logistic model including CTscore, SUVmax, and MI information for predicting the malignancy P was superior to CT and 18F-FDG alone, thus improving physician diagnostic performance.
