Semi-automatic segmentation from intrinsically-registered 18F-FDG–PET/MRI for treatment response assessment in a breast cancer cohort: comparison to manual DCE–MRI

Breast cancer is the most frequent type of cancer in women worldwide [1], with a mean 5-year survival of 90.4% in Norway [2]. Patients diagnosed with locally advanced breast cancer (LABC, stage 3), have a worse survival outcome (78.3%) [2]. They receive neoadjuvant chemotherapy treatment before surgery with the goal of complete pathological tumor regression, which correlates with improved survival and a reduced chance of breast cancer recurrence [3]. Objective response evaluation during neoadjuvant therapy is important assess treatment efficacy and to avoid unnecessary toxic side effects [4]. Radiologically, response evaluation has traditionally focused on measurements of tumor size [5], but several recent studies [6‐10] have established functional imaging modalities as useful indicators of early response during neoadjuvant chemotherapy.

Diffusion-weighted magnetic resonance imaging (DWI) is a functional imaging modality with contrast arising from water molecule motion, and is, therefore, sensitized to tissue microstructure characteristics. DWI is most commonly utilized to assess tissue cellularity, where highly cellular tissues such as malignant tumors exhibit decreased diffusivity [11], quantified by calculation of an apparent diffusion coefficient (ADC). A robust empirical biomarker that is reduced in malignant tumors [12], ADC has shown higher specificity than conventional anatomical MRI for discriminating malignant and benign breast tumors [13].

Tumor ADC is commonly measured by the mean value of manually placed regions-of-interests (ROIs). There is no standard protocol for this tumor segmentation, and different approaches can significantly influence resulting ADC values [14]. Given that direct tumor segmentation of DWI may be confounded by noise and lack of conspicuity, tumor ROIs are commonly delineated on dynamic contrast-enhanced (DCE) images before being transferred to DWI. The definition of tumor on DCE images is thus governed by leakage of gadolinium contrast through pathological vessels and, therefore, linked to vascularity, whereas diffusion changes, reflecting cellularity, do not necessarily coincide [15].

Simultaneous positron emission tomography and magnetic resonance imaging (PET/MRI) is a recent technology with a significant potential in many aspects of breast cancer practice, including diagnostics, staging, and neoadjuvant response evaluation [16]. PET/MRI examinations allow simultaneous collection of structural, functional, and metabolic imaging properties in the same spatial and temporal domain. 18F-fluorodeoxyglucose (FDG)–PET visualizes upregulated glucose metabolism, while MRI reflects other hallmarks of cancer [17] including invasion and metastatic propensity (by ADC) and increased angiogenesis (DCE). Several studies report correlations between standardized uptake values (SUV) from FDG–PET and ADC in malignant tissue [18‐20], indicating that intrinsically-registered 18F-FDG uptake may provide an alternative approach to manually drawn DCE–ROI delineation for use in DWI analysis [19]. FDG–PET is also known to outperform MRI tumor volume measurements in some cancers [21]. In this study, a simple, semi-automated Gaussian mixture model (GMM) segmentation algorithm was selected, to allow for heterogeneous FDG uptake across tumors and expected decline through treatment [22, 23].

The aim of the current study is to investigate the reliability of deriving lesion diffusion imaging characteristics from 18F-FDG uptake in invasive breast cancers > 4 cm or LABC (i.e., cT2-4N0-3) during neoadjuvant treatment. Specifically, we tested the reliability of deriving functional tumor area and ADC values in diffusion-weighted images from intrinsically-registered 18F-FDG–PET uptake using a semi-automated GMM segmentation algorithm in comparison with metrics derived from manually drawn DCE–ROIs.

The main finding of the current breast cancer study is that functional tumor area and corresponding mean ADC values from GMM–PET ROIs matched those derived from manual DCE. As a superficial interpretation, these findings indicate that GMM–PET is a reliable technique to efficiently derive functional diffusion parameters for monitoring neoadjuvant treatment response in breast cancer. The segmentation is data driven, requiring minimal user input, and obviates the requirement for gadolinium contrast administration and, therefore, could have significant potential as an alternative objective evaluation method among the increasing number of breast cancer patients undergoing neoadjuvant treatment. At a deeper level, the results illuminate the concordance—and discordance—between ROIs derived from different imaging modalities, and as such allow interrogation of the spatial relationship existing between functional information arising from PET, diffusion, and DCE imaging, and ultimately the tissue characteristics these modalities are sensitized to.

Our study demonstrates a strong correlation between tumor ADC values derived from GMM–PET and manual DCE segmentation, in line with a previous study by Byun et al. [19] using a similar approach in breast carcinomas; our study utilizes the intrinsic voxel correspondence of simultaneous PET/MRI, thus avoiding the additional registration required by sequential FDG–PET/CT and DWI and conferring greater confidence in the results. Notably, the calculated mean ADC from GMM–PET was not significantly different from mean ADC from manual DCE, despite ADC metrics having been shown to be significantly influenced by segmentation method [14, 31]. This suggests that GMM–PET may have value even while accurate assessment of ADC metrics is considered increasingly important in a neoadjuvant treatment response setting [6‐8].

Conventional manual DCE segmentation means that diffusion measurements, reflecting cellularity [13, 32, 33], are drawn from areas defined by gadolinium contrast enhancement, which is not necessarily optimal and may introduce bias to functional biomarker measurements [15]. It can thus be argued that tumor definition for diffusion studies is better performed on another MR modality more closely related to cellularity. Several studies have described an underlying link between metabolism and cellularity, such as correlation between FDG uptake to cellularity [34, 35]. Consequently, a negative correlation should be expected between SUV and ADC, and it could be argued that GMM–PET would coincide better with changes in cellularity. However, the previous reports are contradictory with either negative [18‐20] or no [36, 37] correlations between SUV and ADC, indicating that imaging metrics from DCE, DWI, and PET do not capture all relevant physiological properties, even when GMM–PET is able to localize tumors equivalent to DCE.

GMM–PET segmentation gives good spatial concordance with manual DCE for untreated breast cancer lesions, while the segmentation performance was significantly poorer for treated lesions, with lower Dice score and higher CoG. It is well known that therapy affects tumor vascularity [38], which may have influenced both manual DCE and GMM–PET segmentation, as both modalities are dependent on sufficient blood flow for contrast and tracer uptake. However, DCE and PET reflect different physiological properties, and therefore, it would be of interest to observe if these have been altered differently during treatment. In addition, it should be noted that several treated lesions with poor overlap are cases where manual DCE segmentation was difficult due to low enhancement on DCE.

In this study, GMM–PET segmentation performed significantly better than the commonly used SUV_42% threshold in untreated lesions, and SUV_2.5 for both treated and untreated lesions in recapitulating manual DCE. Using GMM–PET, the tumor area is not underestimated in the cases of heterogenous uptake (Fig. 5) or for tumors with high-intensity relative to their immediate surroundings, which is a well-known issue [39, 40]. However, as FDG uptake in the tumor decreases during treatment [23], GMM–PET did not perform better than SUV_42%, which is known to give larger estimates of tumor size as SUV_max approaches background levels (Fig. 5), which is a limit of adaptive and data-driven algorithms that are sensitive to the FDG-uptake range [41, 42]. Other algorithm-based approaches such as gradient methods [40, 43] have also been recommended [22], although these require increased user input and were not investigated in this study (Fig. 6).

While the current findings suggest that GMM–PET segmentation can work as a proxy for manual DCE, we do not suggest that PET might replace manual DCE in today’s clinical context. Manual DCE is currently the most sensitive test for both breast cancer detection [44] and monitoring treatment response in a neoadjuvant setting [45], although concerns regarding contrast allergy and potential brain deposition [46] of gadolinium create a setting for exploration of complementary techniques. The use of PET tracer comes with its own challenges, in handling and cost, and is neither available nor suitable in all contexts. Our study indicates that the use of PET data for tumor segmentation is more reliable in pre-treatment lesions; in cases where FDG uptake is substantially reduced by treatment the GMM–PET method becomes less effective, where the tumor may become more diffuse. In these cases, automated segmentation procedures will be more prone to variation. It is worth noting that this is not unique to the technique in this study; the reduction of DCE contrast in successfully treated tumors also makes tumor definition more challenging for the conventional approach. Thus, in a simple sense, PET data are able to act as a proxy for tumor definition using DCE, but are also more likely to become useful as an adjunct to DCE for deeper investigations of tumor characteristics in multimodality examinations. It would also be possible to examine other modality combinations, such as taking a PET ROI definition to derive quantitative DCE markers from a suitable protocol.

Benefits of a data-driven, semi-automated GMM–PET approach include reduced radiologist workload, faster segmentation processes, and reduced interobserver variability [47]. The major advantage of PET/MRI systems, exploited here, is intrinsic registration of simultaneously acquired PET and MR data, giving voxel-level correspondence, without additional registration processes that generally involve user input or additional processing [48]. Further optimization of GMM–PET might include automated data-driven initialization and number of classes, with the goal of entirely eliminating manual input [49]. We illustrate the influence of lesion-region initialization, for which automation could be developed, in Appendix (Fig. 8). The number of voxels within the lesion is also a factor that will affect the performance of the GMM method, and is thus intrinsically tied to image resolution, and will ultimately limit use in smaller lesions.

One limitation to this methodology is that increased tracer uptake is less specific to disease than manual DCE [50], potentially introducing errors in the GMM–PET classification if not accounted for by appropriate selection of tumor-containing region. In addition, while the analysis in this study explicitly did not consider patient response from the clinical viewpoint, as well as other clinically relevant data such as hormone status, the variety of treatments and responses provided a suitably large range of situations in which to test the GMM–PET methodology, including cases where response included an almost total loss of detectable disease.

The potential implications of improved imaging technology in breast cancer are large, and PET/MRI is a unique tool to investigate links between increased metabolism (PET), perfusion (DCE), and decreased diffusion (DWI), without additional scan time or registration errors. The current findings show that PET/MRI, using a semi-automated GMM segmentation strategy, yields tumor area and mean ADC value estimates that can replicate today’s gold standard of tumor definition of manual DCE from MRI. Furthermore, the GMM–PET method also captures tumor changes associated with response to neoadjuvant chemotherapy, which supplements today´s gold standard which is manual DCE in the neoadjuvant setting. The potential benefits include a broader assessment of morphological and metabolic changes to guide clinical decisions regarding tumor operability, and thus to ensure a high probability of complete tumor regression, and subsequent cancer cure.