Computed tomography volumetry of esophageal cancer - the role of semiautomatic assessment

Despite the overall dismal prognosis of patients with esophageal cancer, therapeutic progress has been made and improvement in effectiveness of therapeutic regimens is emerging [1‐3]. At the time of diagnosis, patients with carcinoma of the esophagus often have a locally advanced disease stage with or without distant metastasis [1]. The proportion of patients who can be offered treatment with curative intent is often centered around 25%, a figure which has remained quite stable over time [1, 4‐6]. The predominant symptom generated by these tumors is dysphagia and weight loss. Depending on a variety of factors, the obstruction to the passage of food, through the expanding and stricturing tumor area, results in clinically overt symptoms first at a relatively advanced local stage of the disease [7]. In the evaluation of these patients, accurate staging is mandatory and hereby endoscopic ultrasonography, computed tomography (CT) and positron emission tomography (PET) with fluorine 18 fluorodeoxyglucose (FDG) have taken a central role. The main problem with endoscopic ultrasonography is the dependency on the investigator’s level of expertise [8, 9]. Although FDG PET is frequently used in clinical practice, the scientific validity of this technology has to be better defined [10, 11]. Accordingly, in many referral centers, CT remains the investigation of choice, not only for staging but also for the evaluation of the effectiveness of neoadjuvant therapies [12]. In the attempt to describe the extent of the local tumor growth and also when exploring an eventual therapeutic effect of preoperative therapies, assessment of the volume of the tumor might be critical [12, 13]. Attempts have been made to apply this technique both in controlled as well as uncontrolled research protocols [13‐16]. Some studies indicate that CT-determined volume of esophageal cancer may add to the assessment of neoadjuvant chemoradiotherapy effects and even add prognostic information [13, 17]. However, at present, there is no established and validated method to monitor esophageal tumor response to treatment.

The aim of the current study was therefore to compare the reproducibility of CT volumetry of esophageal tumors using traditional manual segmentation with more modern semiautomatic segmentation by consultant radiologists and radiologists under training.

All tumors were detected by both observers for all included patients. Mean tumor volume when merging arterial and portovenous measurements was 46 ml (range 5-137 ml) using manual segmentation and 42 ml (range 3-111 ml) using semiautomatic segmentation, (p = 0.30). No significant differences in volume were observed between adenocarcinoma and squamous cell carcinoma. All measured volumes are shown in Table 2.

In comparison to manual segmentation, the use of semiautomatic segmentation resulted in a higher interobserver agreement and a lower average absolute percentage difference from mean volume when comparing esophageal tumors volumes segmented by consultant and resident radiologists.

The clinical and research values of CT volumetry at esophageal cancer management are controversial [8, 20, 21]. One possible reason behind this is the fact that CT technology and volumetry techniques used have not been sufficiently addressed and therefore may be suboptimal. For instance, small difference in image contrast between the tumor and the normal esophagus tissues may result in substantial variability in the final calculations. This variation is probably user and experience dependent. In this study, it was observed that the resident, but not the experienced radiologist, had a greater variation when manually delineating adenocarcinoma than squamous carcinoma. This might be explained by small differences in tumor texture [22], which are probably too small to allow the radiologist to diagnose tumor subtype. However, we have recently shown that computerized image analysis, so called CT morphometry, can distinguish between esophageal adenocarcinoma and squamous cell carcinoma [22]. Those small differences in texture might aid the experienced radiologist to better delineate adenocarcinoma but might be too small to the resident to discern. This would explain why the observed difference in tumor volume assessment variation between the resident and consultant radiologist differed between tumor subtype.

One study reported great variations in repeated measurements done by the same observer and also between expert radiologists using the manual segmentation approach [23]. In our study, we observed that by using a semiautomatic segmentation technique, we could significantly reduce this variation to the level of excellent agreement and making the measurements independent on the level of experience of the assessor. A recent study comparing different semiautomatic segmentation software has shown similar excellent intra- and inter observer agreement [24].

Different criteria have been used for the morphological evaluation of esophageal tumors, ranging from bi-dimensional measurement of tumor lesions according to the WHO criteria [25] to the thickness of the esophageal wall [20, 26] or assessment of the volume by use of stereology [27]. In our study, we used the summation-of-area method described by Breiman [28]. This is a simple method which does not require any sophisticated mathematical formulas and has been mainly used to assess the volume of solid organs such as the liver and spleen and also tumor masses e.g. head and neck and kidney.

The observed discrepancy between the readers when evaluating individual tumors can be attributed to several contributing factors. The main individual factor was probably the small difference in image contrast between the tumor and the normal esophagus tissue. This resulted in difficulties in defining the respective cranial and caudal borders of the tumors, especially in tumors located close to or at the gastro-esophageal junction. In addition, the current CT scans consisted only of axial images whereupon no multiplanar reformations were available to the readers. Thinner collimation and coronal and sagittal reformations may add to a better delineation of the cranial and caudal borders of the tumors. Other possible ways to improve the definition of the tumor borders can be to use positive or negative oral contrast media just prior to the CT examination and by the aid of antispasmodic agents [14, 29]. On the other hand, the introduction of a specific, more complex CT protocol for the study of the esophagus might be difficult to implement into clinical routine practice outside tertiary referral centers, where the esophageal tumors are examined with standard CT examinations in the N and M-staging process.

Other methods than CT can add to the armamentarium of methods allowing tumor volume detection and assessment of changes therein. PET-CT imaging with 18 FDG has recently been shown to offer advantages in monitoring the response to neoadjuvant treatment of esophageal cancer by measuring the metabolic/volume activity [21, 30, 31]. However, uncertainties regarding which thresholds of standardized uptake value (SUV) during the delineation of tumor remain as a source of variability in previous studies and there is currently no standardized protocol in use. Analysis of apparent diffusion coefficient (ADC) using diffusion-weighted magnetic resonance imaging (DWI-MRI) has been shown to correlate with histological tumor response and tumor staging [32‐34]. The need to segment tumor volume in order to calculate ADC highlights the importance of reducing the interobserver variation of the tumor segmentation.

Recent developments of computing power have enabled quantification of textural parameters of tumor volumes segmented from both CT and PET images, which has been shown to correlate with overall survival and treatment response in several studies [35‐37], but not in all [22]. However, these methods are sensitive to segmentation errors and accurate segmentation methods are needed in order to ensure comparable results between studies [38, 39].

There are some limitations burdening this study. The number of patients was relatively small and there were only two readers, which exposes the outcome to the risk of the random effect of single outliers. A further sub analysis of differences between segmentation methods depending on histological type might have not shown significance due to lack of enough patients per group (n = 15 versus n = 8). The patients were also scanned on two different scanners. However, this should not impact the comparison between semiautomatic and manual segmentation.

In conclusion, when compared to manual segmentation, application of semiautomatic CT volumetry of esophageal tumors obtained by using modern CT technology, reduces the interobserver variability, regardless of the observer’s experience.