PET/CT guided tuberculosis treatment shortening: a randomized trial

Abstract

Six months of chemotherapy using current agents is standard of care for pulmonary, drug-sensitive tuberculosis (TB), even though some are believed to be cured more rapidly and others require longer therapy. Understanding what factors determine the length of treatment required for durable cure in individual patients would allow individualization of treatment durations, provide better clinical tools to determine the of appropriate duration of new regimens, as well as reduce the cost of large Phase III studies to determine the optimal combinations to use in TB control programs. We conducted a randomized clinical trial in South Africa and China that recruited 704 participants with newly diagnosed, drug-sensitive pulmonary tuberculosis and stratified them based on radiographic disease characteristics as assessed by FDG PET/CT scan readers. Participants with less extensive disease (N=273) were randomly assigned to complete therapy after four months or continue receiving treatment for six months. Amongst participants who received four months of therapy, 17 of 141 (12.1%) experienced unfavorable outcomes compared to only 2 of 132 (1.5%) who completed six months of treatment (treatment success 98.4% in B, 86.7% in C (difference -11.7%, 95% CI, -18.2%, -5.3%)). In the non-randomized arm that included participants with more extensive disease, only 8 of 248 (3.2%) experienced unfavorable outcomes. Total cavity volume and total lesion glycolysis at week 16 were significantly associated with risk of unfavorable outcome in the randomized participants. Based on PET/CT scans at TB recurrence, bacteriological relapses (confirmed by whole genome sequencing) predominantly occurred in the same active cavities originally present at baseline. Automated segmentation of the serial PET/CT scans was later performed, and machine-learning was used to classify participants according to their likelihood of relapse, allowing the development of predictive models with good performance based on CT, PET, microbiological and clinical characteristics. These results open the possibility for more efficient studies of future TB treatment regimens.

Figure 1.

CONSORT recruitment flow diagram. Of 972 participants screened, 704 were enrolled and 146 were withdrawn from the study for the reasons shown (Details of reasons for ineligibility are given in Supplemental Table 1). Of these, 248 were stratified to the non-randomized arm (A) due to the disease severity and 276 were randomly assigned to Arm B or Arm C. *After the DSMB stopped randomization an additional 32 (11 RSA, 21 CHN) participants in the Arm B/C pool were assigned to Arm B (not shown in figure).

Figure 2.

Final study outcomes. A. Kaplan-Meier plot showing the cumulative proportion of unfavorable outcomes starting at week 16, the time of randomization. B. Table showing participant final outcomes by Arm. C. Table of unfavorable outcomes with genomic information, interpretation and whether the relapse involved a cavity present at baseline. *The week 16 isolate was MDR, which may have been from a mixed infection initially, **Subject had a stroke and was not able to be scanned, ***A14 and A15 were relatives and cohabitated during the study, re-infection with the same strain cannot be definitively ruled out (Supplemental Figure 2), ****relapse in lower lobe adjacent to cavity that infiltrated the fissure, *****WGS data and final scan not available for these Chinese participants...

Figure 3.

TB relapse involves baseline cavities. A. Subject A19 presented with a large left apical cavity which appeared mostly resolved after four months of treatment. Six months after completing therapy the participant became sputum culture positive again with the same isolate and with a large left apical cavity in the same region of the lung. B. Participant A9 presented with bilateral disease with a left upper “dirty” cavity with cold central necrotic material and only slight liquefaction. After four months of therapy the right apical cavity showed substantial response while the cold left side lesion appeared less active. 14 months later the participant experienced relapse and became sputum culture positive with an identical isolate and the necrotic left apical lesion appeared to be an active thick-walled cavity.

Figure 4.

Segmentation and computational scan analysis. (A) Axial, (B) Sagittal, (C) Coronal views of PET/CT scan of subject A1. CT scans viewed in lung window and PET at an SUV of 7 are displayed. (D) CT slice from A1 showing cavity and associated features in a single slice, (E) segmented mask for the same CT slice as shown in D, low density normal lung voxels colored red, high density mediastinal and surface of thoracic cavity shown in green, trachea colored yellow, pulmonary vasculature shown in light blue and TB-associated lesional material shown in grey, (F) 3D multiplanar view of the same scan showing the CT bone density colored in green superimposed on a rendering of the 3D masks (low density normal lung removed for clarity) vasculature mask and major airways shown in blue, lesion material shown in peach. (G) 3D rendering of the vascular and lesion masks of A1 colored as in F with the primary CT data including the bones removed. (H) 3D rendering of only the lesion material in A1 with enhanced transparency to view the large left apical cavity. (I) Machine learning model for relapse prediction using Arm C PET/CT scans as the training set and Arms A and B PET/CT scans as the test set. Significant features are Xpert Ct (Baseline): GeneXpert cycle threshold value at baseline, Lung Dissimilarity (HU) (D1): Dissimilarity in Hounsfield Units (HU) between adjacent voxels at a distance of 1 voxel, Lesion entropy (SUV) (D1): Entropy calculated across a lesion at a distance of 1 voxel, Lesion GLCM (gray level co-occurrence matrix) (SD)(SUV)(D1): Standard deviation of the Standardized Uptake Value (SUV) calculated across the lesion at a distance of 1 voxel, Lesion (SD) (HU): standard deviation of HU within a lesion, Lesion (MED40) (HU): Lesion level 40^th percentile of HU value, Lesion Energy (HU) (D1): Energy computed on HU values across a lesion at a distance of 1 voxel, BMI: Body Mass Index.