New strategy for automatic tumor segmentation by adaptive thresholding on PET/CT images

Abstract

Tumor delineation is a critical aspect in radiotherapy treatment planning and is usually performed with the anatomical images of a computed tomography (CT) scan. For non-small cell lung cancer, it has been recommended to use functional positron emission tomography (PET) images to take into account the biological target characteristics. However, today, there is no satisfactory segmentation technique for PET images in clinical applications. In the present study, a solution to this problem is proposed. The development of the segmentation technique is based on the threshold's adjustment directly from patients, rather than from phantoms. To this end, two references were chosen: measurements performed on CT images of the selected lesions, and histological measurements of surgically removed tumors. The inclusion and exclusion criteria were chosen to produce references that are assumed to have measured tumor sizes equal to the true in vivo tumor sizes. In total, for the two references, 65 lung lesions of 54 patients referred for FDG-PET/CT exams were selected. For validation, measurements of segmented lesions on PET images using this technique were also compared to CT and histological measurements. For lesions greater than 20 mm, our segmentation technique showed a good estimation of histological measurements (mean difference between measured and calculated data equal to -0.8 ± 9.0%) and an acceptable estimation of CT measurements. For lesions smaller than or equal to 20 mm, the method showed disagreement with the measurements derived from histological or CT data. This novel segmentation technique shows high accuracy for the lesions with largest axes between 2 and 4.5 cm. However, it does not correctly evaluate smaller lesions, likely due to the partial volume effect and/or respiratory motions.

Figure 1. Reference (CT images): correlation between %Tmean and Tmean/Bmin (a), correlation between %Tmean and Tmax/Bmin (b), and correlation between Tmax and Tmean (c).

Figure 2. Function #4 in three dimensions (a); difference in percentage (%) between functions #4 and #8 (b); difference (in %) between functions #4 and #8, as well as the distribution of “clinical point” (c). (For a better visualization of the “clinical point” localization, their difference value is imposed to ‐3.)

Figure 3. CT image (from FDG PET/CT exam) for lesion #62 (a). FDG PET image (Bmin=464) (b). Zoom on FDG PET image, maximum intensity in the lesion appears in red (Tmean=Tmax=22146), first iteration (c) — FDG lesion uptake in red (%Tmean=52.88% and Threshold=11711, that gives Tmean=17385); second iteration (d) — FDG lesion uptake in red (%Tmean=55.92% and Threshold=9722, that gives Tmean=16214). Convergence at the third iteration (e) (%Tmean=56.83% and Threshold=9215) to a constant lesion area on the PET image (Area PET=62mm2).

Figure 4. Heterogeneity estimation and Area PET (outlined in yellow) for 6 tumors: COV=0.15 for lesion #64 (a); COV=0.20 for lesion #58 (b); COV=0.24 for lesion #63 (c); COV=0.35 for lesion #61 (d); COV=0.36 for lesion #52 (e); COV=0.37 for lesion #56 (f).

Figure 5. Segmentation of lesion #51: on CT image (a), and on the FDG PET image (in red) (b) where the segment outlined in yellow represents the area segmented on CT image (a).