Semi-Automated Co-Segmentation of Tumor Volume Using Multimodality PET-CT in Non-Small Cell Lung Cancer (NSCLC)

Research Article

Austin J Cancer Clin Res 2014;1(3): 1013.

Semi-Automated Co-Segmentation of Tumor Volume Using Multimodality Pet-Ct in Non-Small Cell Lung Cancer (NSCLC)

Li H1, Bai J2, Wu X2, Bhatia SK1, Abu-Hejleh T3, Sun W1, TenNapel MJ1, Menda Y4, Mart CJ1, McGuire SM1, Flynn RF1, Buatti JM1 and Kim Y1*

1Department of Radiation Oncology, University of Iowa, USA

2Department of Electrical and Computer Engineering, University of Iowa, USA

3Department of Internal Medicine, University of Iowa, USA

4Department of Radiology, Division of Nuclear Medicine, University of Iowa, USA

*Corresponding author: Yusung Kim, Assistant Professor, Department of Radiation Oncology, Carver College of Medicine, University of Iowa, 200 Hawkins Drive, 01607PFP-W, Iowa City, IA52242, USA

Received: April 10, 2014; Accepted: May 24, 2014; Published: May 28, 2014

Abstract

Introduction: Lung cancer is the most commonly occurring cancer for both men and women with the highest associated mortality rate. Positron emission tomography and computed tomography (PET-CT) are accurate evaluation modalities in determining lung cancer extent and aggression. An efficient and reliable computerized tumor volume (TV) delineation system, based on PET-CT maging is needed for accurate tumor response analysis during daily clinical practice and in large clinical trial.

Purpose: To present and validate a novel computer-aided, semi-automatic co-segmentation method for lung tumor volume (co-segmented TVPET-CT) that integrates tumor boundaries on both PET and CT.

Methods and Materials: Eighteen patients were included all of whom had stage III/IV NSCLC had received chemoradiotherapy, PET-CT simulation preradiotherapy and a CT scan at 2-4 months follow-up post-radiotherapy. Preradiotherapy GTV (pre-GTV) on PET-CT images were retrospectively contoured by two physicians who reached consensus on the volume and then used this as the reference tumor volume. The statistical correlation were analyzed between the reference tumor volume and different segmented tumor volumes; cosegmented TVPET-CT, segmented Tumor Volume on CT alone (TVCT), segmented Tumor Volume on PET alone (TVPET), Tumor Volume (TVSUV2.5) delineated using the SUV2.5 threshold, and Tumor Volume (TVSUVmean) using the SUVMEAN threshold.

Results: The co-segmented TVPET-CT showed the most significant correlation with pre-GTV (correlation coefficient = 0.993), along with the most favorable ASSD (1.48 ± 0.8 mm) and DSC values (0.85 ± 0.07). TVSUV2.5, TVCT, TVPET, and post-GTV significantly correlated with pre-GTV (P < 0.001) except for TVSUVmean (P = 0.42). Smoking status and histology presented significant correlationwith %ΔTV (P = 0.0273 and P = 0.0297 respectively) while no significant correlations in gender, age, and stage were evident. None of other computeraided segmented tumor volumes or SUVs correlated with %ΔTV. The overall averaged tumor response rate after chemoradiotherapy in 2 - 4 months is 70 ± 18.6%.

Conclusion: The co-segmented tumor volume on PET-CT was most strongly correlated with two-physician-consensus manual contouring that clinically incorporates PET and CT information.

Keywords: PET-CT; Computer-aided segmentation co-segmentation; Non-small cell lung cancer; PET

Introduction

Lung cancer is the leading cause of cancer-related death in the world [1]. During 2013 it caused 159,260 deaths in the United States [2] with only a 16.3% 5-year overall survival [3]. Eighty-five percent of lung cancers are diagnosed as NSCLC, and 70% of patients present with advanced disease (stage III - IV) [4,5]. Clinical decision making is generally predicted upon tumor response with complete response defined as tumor disappearance, partial response as more than 25% decrease in size measured by CT, stable disease as less than 25% decrease or increase, progressive disease as greater than 25% increase in the tumor diameter using RECIST criteria (response evaluation criteria in solid tumors) [6,7]. Consequently, there is a compelling need for the more accurate measurement of the tumor volume for identifying tumor extent for treatment and quantifying tumor response to treatment.

Studies suggest that the combination of functional (PET) and anatomical imaging (CT) (PET-CT) is a reliable indicator of tumor extent and distribution [1,8]. The use of PET-CT imaging alteredthe tumor volume delineation of NSCLC in more than 50% of patients when compared with CT-based tumor volume alone [9,10]. Information obtained from combined PET-CT imaging that cannot be obtained from PET only data for lung cancer patients enables the physician to differentiate between anatomic pathological and physiological changes. In radiation oncology practice, manual contouring is still needed in order to delineate the tumor on PET-CT in order to reconcile the two tumor volumes (TVs) defined separately on CT and PET; a process that is time consuming, cumbersome, and error-prone. It also suffers from substantial inter- and intra-observer variability, which limits its utility in large-scale clinical trial research [11]. Reliable auto-segmentation methods that quantitatively analyze PET-CT datasets quickly, robustly, and objectively are presently not clinically available. Existing PET-CT segmentation methods currently in use either work only for a single modality (PET or CT) or work on one image set represented by the fused PET-CT datasets [12].

In this study we propose a novel co-segmentation framework using PET-CT, where TV is co-segmented simultaneously, yet separately, using both PET and CT datasets while admitting the uncertainties described above.

Methods and Materials

PET-CT Co-segmentation algorithm

The process for co-segmenting a tumor volume on PET-CT (cosegmented TVPET-CT) starts from a physician-identified tumor location (Figure 1A). This is accomplished by asking the physician to identify the center of the tumor using three planes on the PET-CT. This is the only step requiring a clinician’s input. Afterwards, segmentation is simultaneously performed on each CT and PET image, and globally optimized using the mutual information from both segmentations to generate a co-segmented TVPET-CT based on the algorithm described by Song et al. [13]. To co-segment the tumor from both PET and CT scans, we added a PET-CT context term EPET-CT to the energy functions of CT and PET (Equations (2) and (3) in Ref [13]), which penalizes the segmentation difference between the two image datasets (Figure 1 B). Without loss of generality, we assume that the PET and CT images, IP and IC, are registered. Let (p, p’) denote a pair of corresponding voxels in IP and IC. We penalize the label difference with δ pp' ,( f p P , f p C ) [email protected]@[email protected]@+=feaaguart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLnhiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaGaeqiTdq2aaSbaaSqaaiaadchacaWGWbGaai4jaaqabaGccaGGSaGaaiikaiaadAgadaqhaaWcbaGaamiCaaqaaiaa[email protected][email protected] for p and p’. The PET-CT context term then takes the form:

E PETCT ( f P , f C )= ( p, p ' )withp I P ,p' I C δ p p ' ( f p P , f p ' C ) [email protected]@[email protected]@+=feaaguart1ev2aqatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLnhiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8qacaWGfbWdamaaBaaaleaapeGaamiuaiaadweacaWGubGaeyOeI0Iaam4qaiaadsfaa8aabeaak8qadaqadaWdaeaapeGaamOza8aadaahaaWcbeqaa8qacaWGqbaaaOGaaiilaiaadAgapaWaaWbaaSqabeaapeGaam4qaaaaaOGaayjkaiaawMcaaiabg2da9maawafabeWcpaqaa8qadaqadaWdaeaapeGaamiCaiaacYcacaWGWbWdamaaCaaameqabaWdbiaabEcaaaaaliaawIcacaGLPaaacaWG3bGaamyAaiaadshacaWGObGaamiCaiabgIGiolaadMeapaWaaSbaaWqaa8qacaWGqbaapaqabaWcpeGaaiilaiaadchacaqGNaGaeyicI4Saamysa8aadaWgaaadbaWdbiaadoeaa8aabeaaaSWdbeqan8aabaWdbiabggHiLdaakiabes7aK9aadaWgaaWcbaWdbiaadchacaWGWbWdamaaCaaameqabaWdbiaabEcaaaaal8aabeaak8qadaqadaWdaeaapeGaamOza8aadaqhaaWcbaWdbiaadchaa8aabaWdbiaadcfaaaGccaGGSaGaamOza8aadaqhaaWcbaWdbiaadchapaWaa[email protected][email protected]

Note that the PET-CT context constraint is soft with δ ( f p P , f p C ) p p <+ [email protected]@[email protected]@+=feaaguart1ev2aqatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLnhiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8qacqaH0oazdaWgbaWcbaGaamiCaiqadchagaqbaaqabaGccaGGOaGaamOza8aadaqhaaWcbaWdbiaadchaa8aabaWdbiaadcfaaaGccaGGSaGaamOza8aadaqhaaWcbaWdbiqa[email protected][email protected] , which accounts for the tumor boundary differences between PET and CT. The corresponding voxels in PET and in CT could be assigned different labels (“object” or “background”) if prominent features present very differently in PET from in CT, which could be caused by imaging uncertainties, registration errors, or both. Our method is thus able to accommodate those uncertainties, further improving its applicability. The energy function of our co-segmentation algorithm is defined as follows:

E cs ( f P , f C )= E PET ( f P )+ E CT ( f C )+ E PETCT ( f P , f C )       (1) [email protected]@[email protected]@+=feaaguart1ev2aqatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLnhiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=xfr=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[email protected][email protected]

We used a more flexible PET-CT context term EPET-CT to make use of the dual modality information [13]. We also minimized the cosegmentation energy function Ecs by solving a minimum cost s-t cut problem in a transformed graph G, which allows a globally optimal solution in low-order polynomial time. To enforce the PET-CT context term EPET-CT, additional inter-sub-graph arcs are introduced between GP and GC. Figure 1 B illustrates the construction of the graph G, with green arcs encoding the PET-CT context term EPET-CT. The minimum-cost s-t cut in G defines an optimal delineation of tumor volume in both PET and CT images with respect to the energy function (1).