Quantitative 18F-FDG PET-CT scan characteristics correlate with tuberculosis treatment response

Participants considered for this report were 99 newly diagnosed, culture-confirmed, pulmonary TB patients who successfully completed follow-up as part of the previously published Catalysis treatment response cohort [44‐46]. They were HIV-uninfected adults, recruited at primary health care clinics in the northern regions of Cape Town, South Africa. Patients underwent FDG PET-CT scans at diagnosis (Dx) and at month 1 (M1) and month 6 (M6) of treatment. Fifty patients also had FDG PET-CT scans 1 year after the end of treatment (EOT + 1y). PET images were corrected for attenuation and reconstructed to 4 × 4 × 4 mm voxels using an iterative algorithm. The CT scan parameters were set at 120 kV, 100 mAs, without dose modulation with 1.17 × 1.17 mm pixels, and a 3-mm slice thickness, reconstructed with I31 filter and B31 s con kernel.

Figure 1 shows a flow diagram of the study design and scan settings as described in Additional file 1: Supplementary note 1.

Clinical samples and information were collected at day 0, week 1, 4, 8, 12, and 24 (month 6) for all participants. Samples included liquid culture with speciation and GeneXpert® MTB/Rif (Xpert) assays on sputum and the analysis of multiple biomarkers in blood and urine.

We previously conducted and reported qualitative scan assessments by comparing each lesion’s intensity at M6 to the intensity at Dx.⁴⁶ Three different response patterns were described: (1) A ‘resolved’ scan response pattern showed no lesion with more than minimally increased FDG intensity when compared to surrounding lung tissue M6. (2) An ‘improved’ pattern indicates that all lesions improved during treatment, but one or more lesion showed residual FDG avidity at M6. (3) A ‘mixed’ response indicated that while some lesions improved, at least one intensified, or a new lesion was present at M6. EOT + 1y scans were compared to M6 scans for qualitative classification.

Based on a previously described methodology [47], we quantified the extent and severity of lung lesions concurrently on PET and CT for all scans (Dx, M1, M6, EOT + 1y). After co-registration of scans across time points using the SPM toolbox [48] in MATLAB (Mathworks Inc.), we created volumes of interest (VOIs) of lungs on the CT component with MRICro [49]. The lung VOIs were adapted by excluding areas affected by misregistration and created to fit all time points. In some cases, we had to create a separate lung VOI for the EOT + 1y scan, due to substantial lung volume changes related to fibrosis or poor inspiratory effort. In addition, we created VOIs which appeared lesion-free on both PET and CT on all time points to represent references for background FDG uptake in the lung.

We segmented the PET component by using a lesion-to-background comparison. To reduce intra- and inter-scan variability, we standardised uptake using patient-specific reference volumes. We assigned a Z-score to each voxel based on:

in which μ_NL and σ_NL are the mean and standard deviation of PET counts within the lesion-free lung VOIs for each scan. All voxels exceeding a Z-score of 8 were segmented as FDG-avid [47].

We used a previously reported method of density thresholding to segment lesions on CT [42‐47]. These thresholds were as follows: (1) normal density, between − 950 Hounsfield units (HU) and − 500 HU; (2) soft lesions (V_soft), from − 500 to − 300 HU, usually tree-in-bud lesions or nodules, but may also include regular, medium, to large vasculature; (3) medium density lesions (V_medium) from − 300 to − 100 HU, which usually consists of nodular infiltrates, but may also include established lesions in early progression or partial resolution; and (4) hard lesions (V_hard), above − 100HU, are usually due to consolidation, cavity walls, bronchial thickening, or calcified fibrosis. We delineated cavity air volume using a gradient-based region-grow technique, and on the M6 scan measured the thickness of cavity walls at the level of the widest diameter on the transverse view. We measured the wall thickness of enclosed cavities at the level of widest cavity diameter, and the area of maximum wall thickness where there was no confluence with other lesions and structures, or fibrotic changes (examples shown in Additional file 1: Figure S1).

After segmentation, the following PET parameters were quantified: (1) metabolic lesion volume (MLV); (2) the mean Z-score in the MLV (Z_mean); and (3) total glycolytic activity index (TGAI): the product of the MLV and mean lesion- to- background intensity index:

(TGAI = MLV × meanlesioncounts/meancountsinnormallung).

In addition, the program also measured the volumes of each abnormal density category on CT, i.e. V_soft, V_medium, V_hard, and total volume with abnormal density > − 500 HU (V_total). We also measured a combined FDG PET-CT metric: MLV_abN = the intersection of MLV and area with increased density on CT (≥ − 500 HU).

To create a variable to combine all major contributing factors on PET and CT, we assigned the Z_mean score to the cavity volume (with no perfusion, thus no FDG uptake). We then added this to the TGAI value to obtain a composite measure of both metabolically active lesions and cavities. The resulting formula was:

We previously published further detail regarding quantification and evaluation of the described technique [47].

We considered a P value smaller than 0.05 as significant. We tested the association between categorical and continuous variables with a two-way T test for independent samples (Tibco© Staticstica™ V13). For analysis of variance tests with multiple grouping variables (TTN grouping variables), we applied the more conservative Kruskal-Wallis non-parametric test (R version 3.2.2). We calculated P values for receiver operating curves in order to test the null hypothesis that the area under the curve equals 0.50 (GraphPad© Prism V8). We performed post hoc analysis to distinguish favourable and pooled unfavourable outcomes by applying thresholds suggested by receiver operating curves. We used the Fisher exact test to determine significant associations between categorical variables (GraphPad© Prism V8). In this descriptive study, we did not correct for multiple testing.

We use the standard terms prognostic and predictive when comparing the association between FDG PET-CT parameters and outcomes. While the ability of a M6 marker to identify failed cases is strictly speaking diagnostic since it applies to the same time point, in practice, culture results are delayed and failed and relapse cases are grouped together when assessing treatment efficacy.

We recruited 99 PTB patients, of which 95 had drug-sensitive (DS) strains, 2 had isoniazid mono-resistant strains, and 2 multi-drug resistant strains. More details regarding the treatment regimens are provided in Additional file 1: Supplementary note 2.

We based patient clinical outcome classifications on WHO definitions, except we used the more sensitive sputum culture, instead of direct smear microscopy. The outcomes for 3 patients were classified as un-evaluable (UE) due to sputum culture contamination, and they were excluded from the analysis. Of the remaining 96 participants, favourable treatment outcomes include 76 cured cases (achieved and maintained culture conversion). Unfavourable outcomes included 8 failed treatment cases (sputum culture positive at M6), and 12 with recurrent PTB (initially culture converted, but re-diagnosed with active TB within 2 years after treatment completion). One of the 8 failed treatment cases was asymptomatic in spite of a positive sputum culture at M6 and declined to restart treatment and remained symptom-free when assessed a year later. Of the patients with recurrent PTB, 2 were culture confirmed; 5 were confirmed by both Xpert and smear positivity by direct microscopy (acid-fast bacillus positive); 3 were Xpert negative at month 6 but converted back to positive; and 3 remained Xpert positive for more than 6 months and deteriorated clinically. The absence of post-treatment culture complicated the recurrence diagnosis and prevented the distinction between relapse and reinfection.

Outcome was further stratified based on time to culture conversion and treatment adherence. Eighteen patients converted to sputum culture negative within 4 weeks, an additional 39 by week 8, another 22 by week 12, and a further 9 by week 24. Time to culture negativity (TTN) was un-evaluable (UE) for 3 patients due to contaminated cultures. Fourteen patients took fewer than approximately 80% of their treatment dosages during the 6-month period, which is regarded as poor adherence in most clinical trial designs. The failed treatment group included 4 patients with poor treatment adherence and 1 with MDR disease [12, 13]. Further clinical information and demographics of the cohort may be found in our previously published online methods [44].

We performed a fourth scan 1 year after the end of treatment (EOT + 1y) for 50 patients that culture-converted at M6. Eight of these 50 patients were diagnosed with recurrent disease by healthcare providers within 2 years of treatment completion (five before the EOT + 1y scan and three after). The other 42 maintained favourable treatment outcome status.

The scans showed ongoing inflammation at the end of treatment in the majority of the patients [44]. For 51 (52%), there was an improved response on the M6 scan (Fig. 2a, Fig. 3a). A mixed response was seen in 34 (34%) patients (Fig. 2b, Fig. 3b), of which 14 had both new and more intense lesion(s), 16 demonstrated only an increase in the intensity of lesion(s), and 4 had only new FDG-avid lesion(s). Only 14 (14%) patients had a resolved pattern on their M6 scan (Fig. 2c).

The morphology associated with the most intense lesion of each mixed and improved M6 scan included CT features suggestive of active PTB, such as cavities (in 26 cases), patchy consolidation (in 22), complex lesions involving consolidation with cavitation (in 16), nodular infiltrates (in 17), enlarged hilar lymph nodes (in 3), and pleural-based infiltrates (in 1). Smaller nodules and tree-in-bud-lesions without calcification tended to resolve during treatment, especially when present in the lower lobes, and even if they were diffuse. Results for each patient are included in Additional file 2: Dataset 1.

Lesion burden was significantly associated with TTN for the three main independent FDG PET-CT parameters (total cavity volume, TGAI, V_total) at Dx, M1, and M6 (Fig. 4a, c, e). The differentiation between the TTN groups became more pronounced during treatment. Cavity volume showed the largest difference between TTN groups at single time points (P < 0.001 at Dx, M1, and M6 —Fig. 4c). Proportional TGAI changes (P = 0.031 at M1; P = 0.002 at M6 —Fig. 4b) from baseline (delta), were also significantly associated with TTN. Similar trends were noted for cavity volume and V_total (Fig. 4d, f), but did not meet the threshold for significance. Delayed sputum converters (between months 5 and 6) and failed treatment cases thus showed both a larger burden of disease and a slower rate of reduction in scan metrics. The recurrence group also showed a slower rate of reduction in scan metrics, but did not have a large baseline burden of disease (Fig. 4d).

We also evaluated the individual components of the TGAI (Z_mean, the SUV_max, MLV —Additional file 1: Figure S3), and high-density lesions on CT (V_hard, V_medium, V_soft —Additional file 1: Figure S4), as well as the intersection of high-density lesions on CT and FDG-avid lesions on PET (MLV_abn—Additional file 1: Figure S5). We found a significant correlation between TTN groups and single time-point values for indicators of lesion volume (MLV, V_hard, V_medium, V_soft, MLV_abn), but not for indicators of PET intensity (Z_mean, SUVmax). The proportional change from Dx to M6 in these variables was significantly associated with TTN groups for all these variables (Z_mean, MLV, V_hard, V_medium, V_soft, MLV_abn) except SUV_max. None of these variables, however, showed a clear advantage over the main independent FDG PET-CT variables (cavity volume, TGAI, V_total).

As expected from TTN correlation results, cavity volume had the strongest association with treatment failure, with an area under the curve (AUC) of 0.81 (P = 0.006) at Dx, 0.83 (P = 0.005) at M1, and 0.87 (P = 0.004) at M6 (Additional file 1: Figure S6). Apart from cavity volume, other metrics reflecting lesion extent (V_total, MLV, MLV_abn) also showed promise to differentiate failed cases from cured at Dx, M1, and M6, while parameters reflecting the intensity of FDG uptake (SUVmax, Z_mean) did not show prognostic value at baseline, but only at M6. A summary of AUC’s for various scan parameters to differentiate failed treatment cases is in Additional file 1: Table S1. Of note, the one asymptomatic, failed treatment case (participant identification number 43) had quantified values in keeping with a good response to treatment.

In addition to cavity volume, M6 cavity wall thickness was also associated with treatment failure. In most cured cases, M6 cavity wall thickness ranged from 0 (no cavity) to 3 mm. M6 cavity wall thickness in failed cases was significantly greater than cured cases’ (Student’s T test for independent samples; P <  0.001) and ranged from 2.5 to 8 mm. At end of treatment, recurrent cases were not significantly different from other cured cases; their M6 cavity wall thickness ranged from 0 to 4 mm.

The 12 patients diagnosed with recurrent disease within 2 years after treatment had a similar sputum culture conversion rate to cured cases (median TTN 8 weeks) and did not show a comparatively large lesion burden at Dx (Fig. 4). Nevertheless, irrespective of TTN, during treatment they exhibited a relatively slow rate of reduction in TGAI, cavity volume, and to a lesser extent V_total (Fig. 4b, d, f).

At M1, there was a trend for the recurrent disease group to have a smaller reduction in TGAI and TGAI_com burden. At M6, the difference between the groups was significant (P = 0.003). No other parameters were significantly different between cured and recurrent disease groups.

Patients who reported previous PTB episode(s) tended to have a higher TGAI burden at Dx and showed significantly less TGAI reduction on treatment (P = 0.003, Additional file 1: Figure S7). They also showed less reduction in cavity volume, but no clear difference in abnormal CT density (V_total). See Additional file 1: Table S2 for additional summary statistics on previous TB.

We pooled patients with unfavourable outcomes (failed and recurrent treatment) and analysed the most promising scan parameters’ distribution per groups, combined with receiver operating curve analysis to determine the most informed thresholds. A failure to reduce TGAI by less than 80% from Dx to M6 was the scan characteristic most associated with unfavourable outcomes and carried an almost sevenfold risk. Table 1 shows indicators of M6 scan parameter associations with unfavourable outcomes and the suggested cut-offs, and Fig. 5 compares the distribution of scan metrics for outcome groups.

A total M6 cavity volume greater than 7 ml and a M6 TGAI of greater than 600 (equivalent to a SUV-based total glycolytic activity of roughly 200 if calculated using SUV) also carried a fourfold increased risk of an unfavourable outcome. Cavity volume was slightly more sensitive and TGAI more specific in predicting unfavourable outcomes. Combining the variables did not improve prognostic accuracy, either when merged into a single variable or when used in Boolean selection. TGAI_com performed very similarly to TGAI. We also tested whether either M6 cavity volume > 7 ml or TGAI > 600 put a patient in the high-risk group, but this generated the same sensitivity but lower specificity than single variables.

Quantitative parameters at M6 out-performed lesion-based qualitative measurements. A mixed response pattern (either new or intensified lesions) at M6 was associated with a 2.86 times increased risk of unfavourable outcome, which was comparable to the prognostic potential of the quantitative parameters at M1. At M1, both a cavity volume greater than 20 ml and a less than 33% cavity volume reduction from Dx were significantly associated with unfavourable outcome, showing a 2.5- and 2.8-times increased risk of unfavourable outcome respectively (P = 0.04 and P = 0.01 respectively). Total TGAI values at Dx and M1 did not perform well as a predictor of pooled unfavourable outcomes, but a trend (P = 0.07) suggests an increased risk of unfavourable outcome when there is less than 5% reduction in TGAI (from Dx to M1).

Most residual lesions were smaller and less intense 1 year after the end of treatment. Metabolic lesion volume decreased to an average of only 2.03% of lung volume at EOT + 1y, compared to 4.21% at M6 in the same patients. Mean total cavity volume also decreased from 7.6 ml at M6 to 2 ml at EOT + 1y. Abnormal CT density showed less reduction during treatment than other parameters after treatment, and the mean V_total at EOT + 1y was 4.53%, compared to 5.7% at M6. Figure 6 shows the distribution of TGAI and cavity volume across time points.

Remarkably, only 32% of EOT + 1y scans were completely resolved. The remaining 68% had FDG-avid residual lesions, of which half had improvement of all lesions compared to M6 (Fig. 2a), and the other half had a mixed lesion response compared to the M6 scan (Fig. 2b, c, Fig. 3c). Morphology of new FDG-avid lesions at EOT + 1y included nodular infiltrates (found in 4 cases), hilar lymph nodes (in 1), cavitation (in 2), consolidation (in 2), or lesions with combined morphology (in 3). There was no association between the development of new lesions during Dx-M6 and during M6-EOT + 1y. Morphology of residual M6 lesions showing similar or more intense FDG uptake at EOT + 1y included consolidation (2), cavitation (4), and nodules (2). All three patients who developed recurrent PTB after EOT + 1y had mixed scan outcomes at this time point (Fig. 2c), while none with resolved EOT + 1y scans were diagnosed with recurrence. Figure 2 and Fig. 3 show examples of dynamic lesion progression and resolution during and after treatment.

We found no significant association between M6 and EOT + 1y for TGAI or cavity volume (Additional file 1: Figure S8a and S8b). However, we found a moderate correlation between the time points for V_total and SUVmax (Additional file 1: Figure S8c and S8d).

Quantification of the FDG PET-CT images provides metrics that show stronger association with clinical outcomes compared to qualitative scan patterns. Qualitative scan response patterns are more challenging to interpret, due to varying responses of individual lesions and incomplete resolution of inflammation during treatment. The most promising quantitative marker (TGAI not reducing by > 80% from Dx to M6) carried a 6.97 relative risk of unfavourable outcome, compared to 2.86 if a mixed response pattern was observed.

Various scan metrics measured in this study showed prognostic potential at Dx and M1 and stronger associations with unfavourable outcomes by M6. A high cavity volume showed the strongest association with a risk of treatment failure, while a suboptimal reduction of the total glycolytic activity throughout the lung had the strongest association with recurrent disease. Both of these variables also correlated with time to culture negativity. This suggests a correlation between the quantified lesion burden and the MTB load that is clearer when using quantitative rather than qualitative analysis.

The volume of high-density lesions on CT (V_total) also shows a strong association with TTN and failed treatment, even at early time points (Dx and M1). Unlike TGAI, however, V_total shows no association with recurrent disease and subsequently, pooled unfavourable outcomes. This is likely due to residual scarring and fibrotic changes and the related residual abnormal density lesions on CT after treatment. Values indicating FDG uptake intensity alone (SUVmax and Z_mean) do not show association early in treatment. However, the proportional intensity changes are associated with TTN and outcome, though not as strongly as TGAI, which combines information from intensity and volume. Combined PET and CT parameters perform similarly to their underlying components, but they do not appear to be clearly superior to individual variables.

Quantification of EOT + 1y scans confirms our previous observations that there is a tendency for all parameters to decrease after treatment, but that a lack of complete resolution is still common and new or intensifying lesions are often seen. Dynamic changes after treatment are common for PET parameters resulting in a poor correlation between M6 and EOT + 1y measurements, compared to CT lesions which appear to be more persistent after treatment.

We found notable differences between failed and recurrent treatment cases. Failed treatment is associated with extensive lung lesions at baseline and large cavities with thick walls at M6, as well as poor adherence. On treatment, a reduction in the FDG avidity and thickness of cavity walls is usually also associated with a reduction in cavity volume. Interestingly, it is relatively common for cavities to show a reduction in both FDG avidity and wall thickness (thus appearing inactive), but to show an increase in size after M1. This may reflect a loss of structural wall strength and progress towards the formation of bullae (Additional file 1: Figure S1b). Recurrent cases display a comparatively low lesion burden at baseline, average adherence, and time to sputum culture negativity, but insufficient reduction in lesion burden during treatment. We also found insufficient reduction in lesion burden in patients with a history of previous PTB episodes.

The catalysis treatment response cohort is the largest prospective study conducted on the use of FDG PET-CT in human patients with PTB and the first report on the fate of residual FDG PET-CT lesions after PTB treatment [45]. In this report, we found that a quantitative analysis of scan characteristics shows a stronger association with outcomes than a qualitative analysis of these same characteristics. In related publications, these quantitative metrics also correlate well with host biomarkers, namely gene expression signatures [46], and urinary concentration of the recently discovered metabolite, seryl-leucine core 1 O-glycosylated peptide [50]. The potential of quantitative FDG PET-CT variables to identify patients with a low risk of treatment failure was also analysed in combination with other patient variables and is currently being tested in the PredictTB trial [51].

Our quantitative findings correspond well with previous reports on animal models [29‐31, 52] and validate findings from two studies in drug-resistant TB cases, which also show that the quantified inflammation burden, as measured by FDG, corresponds with the effectiveness of treatment [41, 43]. Our results are consistent with those of previous reports in humans in which cavitary disease is associated with an unfavourable outcome [11, 53, 54]. The persistence of density changes in the lungs is also in keeping with reports of the high incidence of post-tuberculosis lung impairment [6, 8]. We found no published reports in human tuberculosis that compare as many quantified parameters in either PET or CT scans for a sample size this large or suggest cut-off values that may be used for direct comparison in future studies.

Although this is the largest prospective cohort of FDG PET-CT in TB treatment response, it is still a limited sample size with a small number of unfavourable outcomes, and we did not have sufficient data to differentiate relapse from recurrence. On account of these limitations, we did not perform multivariable logistic regression analysis, due to both the risk of false-positive findings from overfitting a model and false-negative findings due to confounding unfavourable outcomes by reinfection cases. In the absence of a ground truth, we also cannot exclude the possibility that other analysis methods could improve on the prognostic ability of scan characteristics or that different PET-CT scanner models, acquisition, and reconstruction protocols will affect the results. With regard to EOT + 1y scans, the variation in timing between the scan and the recurrent disease diagnosis limited the conclusions we can draw from the data. The study design excluded HIV-infected participants to ensure a more homogenous group for biomarker discovery. As such, the suitability of the variables and cut-offs will have to be re-evaluated in this important subset of TB patients. We did not perform pharmacokinetic studies.

At month 6, the best indicator of unfavourable outcomes (treatment failure and recurrences) shows 80% sensitivity and 75% specificity, which is modest for a diagnostic test but far superior to currently used predictive biomarkers for poor treatment outcomes, such as month 2 sputum culture conversion or AFB smear conversion. The thresholds defined in this report require further validation before direct application to clinical practice, but the associations with quantified patient and microbiological data suggest that most prominent FDG PET-CT scan characteristics can be used as part of risk stratification and treatment response monitoring in therapeutic trials. In combination with clinical evidence, it may also assist treatment decisions in complicated clinical cases, such as treatment of drug-resistant TB and evaluation of adverse reactions to medication. Quantification of central trends in lesion burden provides continuous variables that allow multiple options for statistical analysis. This approach holds promise for improving the accuracy of clinical reporting.

Quantification methods should be further improved to be less operator-dependent, more user-friendly and widely available to both researchers and clinicians. Suggested thresholds should be validated and tested in shortened TB regimens. Further translational research may implement FDG PET-CT scan characteristics to explore complex interactions between host, MTB, and anti-tuberculous drugs, to help develop improved regimens, host-directed therapy, and diagnostic tests.