Jonathan Verma, MD, Radka Stoyanova, PhD, N. Andres Parra, PhD, Kyle Padgett, PhD, Michael Samuels, MD, Nagy Elsayyad, MD; University of Miami Sylvester Comprehensive Cancer Center
PURPOSE: Recently, an unsupervised pattern recognition (UPR) technique was developed by one of the investigators to analyze dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). The technique utilizes non-negative matrix factorization (NMF) and has hitherto not been studied in head and neck cancers (HNCs). This study aims at implementing this technique in HNCs in order to explore its potential value in characterizing perfusion patterns within these tumors. The presence of subvolumes in the HNC that are differentially perfused may have future therapeutic implications with respect to dose painting, intensification, or deintensification of treatment.
METHODS: Eight patients who received definitive chemoradiotherapy for locoregionally advanced HNC were randomly selected. All received 70 Gy/33–35 fractions/6–7 weeks with concurrent cisplatin. As part of the radiotherapy planning, MRI had been obtained for each patient, and DCE sequences were acquired, in addition to the usual T1 and T2 sequences. DCE-MRI datasets with contoured gross target volumes (GTVs) were imported for image analysis into MIM software (MIM Corporation, Cleveland, Ohio). The NMF (non-negative matrix factorization) algorithm was implemented in MIM. There are three characteristic patterns of signal-vs-time curves that UPR identifies within each GTV—namely, Pattern A, which identifies subvolumes with fast contrast uptake and fast washout; Pattern B for subvolumes with delayed uptake and washout; and Pattern C, with slow or no contrast uptake.
RESULTS: The aforementioned three patterns identified corresponding subvolumes in each GTV. The mean GTV volume was 56.8 cc (range: 22–146.6 cc). Subvolumes for each of Patterns A, B, and C averaged 31% ± 3.6%, 30% ± 5.3%, and 39% ± 8.4% of the total GTV, respectively. The standard deviations showed relatively tight dispersion of these values around their respective means, suggesting relative constancy throughout the patients studied. Although no clear geographically consistent relationship between the subvolumes emerged, necrotic regions identified by computed tomography (CT) or MRI were consistently located within Pattern C subvolumes. All three patterns were qualitatively identical to those reported in a previously published preclinical study by our group, wherein Pattern B subvolumes corresponded to regions of tumor hypoxia on 18F-fluoro-misonidazole (MISO) images and radiolabeled pimonidazole histological slices.
CONCLUSION: UPR is a novel technique that has been developed in a preclinical tumor model and that is potentially applicable in HNCs. This exploratory study appears consistently to identify subvolumes within the GTV that may have therapeutic implications and that may be promising areas for further research (eg, correlation with hypoxia and clinical outcome).
Proceedings of the 97th Annual Meeting of the American Radium Society — americanradiumsociety.org