|Year : 2020 | Volume
| Issue : 2 | Page : 56-61
Computed tomography texture-based radiomics analysis as a predictor of response in head-and-neck cancer treated with chemoradiation and hyperthermia
Nagraj Huilgol, Balaji Ganeshan, Gopal Pemmaraju, Anand Parab, Anuradha Singh
Department of Advanced Centre for Radiation Oncology (ACRO), Nanavati Super Speciality Hospital, Dr. Balabhai Nanavati Hospital, Mumbai, Maharashtra, India
|Date of Submission||07-May-2020|
|Date of Acceptance||08-May-2020|
|Date of Web Publication||23-Jun-2020|
Dr. Gopal Pemmaraju
Department of Advanced Centre for Radiation Oncology (ACRO), Nanavati Super Speciality Hospital, Dr. Balabhai Nanavati Hospital, Mumbai, Maharashtra
Source of Support: None, Conflict of Interest: None
Aim: In this pilot retrospective study, we aim to assess the ability of computed tomography (CT) texture-based radiomics analysis to predict the response in locally advanced head-and-neck cancer patients treated with chemoradiation and hyperthermia. Materials and Methods: Patients with locally advanced head-and-neck cancer who were treated with multiple modalities of chemoradiation and hyperthermia or radiation with hyperthermia between January 2018 and December 2019 were analyzed. The initial response was scored as complete response (CR), partial response, no response, or progressive disease. CT representing central core of the tumor, and 2 cm on either pole was segmented. These segmented sections of CT scans were analyzed on TexRAD software. Results: A total of 25 patients' data were analyzed, of which 15 are males, and most of the patients had Stage III or IV disease. Eleven patients received ≥10 sittings of hyperthermia. Only those who had CR 16 patients were included as responders and the rest as nonresponders (NRs). A higher kurtosis value is associated with non responders. Receiver operating characteristics further demonstrated kurtosis at fine texture scale ≥1.29 suggested NR. A total number of pixels of the center core is a significant differentiator between CR and NR (P = 0.039). Similarly, the mean density of the top section is a significant differentiator (P = 0.044). A higher mean density is indicative of NR. A higher composite score is associated with NR. Conclusion: Texture-based radiomics analysis on CT is an inexpensive, noninvasive addition based on imaging for prognostication potentially acting as an imaging biomarker. A composite score – a new paradigm as illustrated in this article demonstrates promise in assessing the response in head-and-neck cancer patients treated with hyperthermia in addition to radiation or chemotherapy, which needs to be further confirmed in a larger, prospective study.
Keywords: Area under the curve, complete response, computed tomography texture analysis, computed tomography, no response, positron emission tomography, receiver operating characteristics
|How to cite this article:|
Huilgol N, Ganeshan B, Pemmaraju G, Parab A, Singh A. Computed tomography texture-based radiomics analysis as a predictor of response in head-and-neck cancer treated with chemoradiation and hyperthermia. J Radiat Cancer Res 2020;11:56-61
|How to cite this URL:|
Huilgol N, Ganeshan B, Pemmaraju G, Parab A, Singh A. Computed tomography texture-based radiomics analysis as a predictor of response in head-and-neck cancer treated with chemoradiation and hyperthermia. J Radiat Cancer Res [serial online] 2020 [cited 2020 Dec 5];11:56-61. Available from: https://www.journalrcr.org/text.asp?2020/11/2/56/287448
| Introduction|| |
Hyperthermia in conjunction with chemoradiation or radiation alone has shown to be an effective approach to the treatment of locally advanced head-and-neck cancer. Heating tumors ensconced deep within the body or even in the head-and-neck area is a challenging task. Response to heat depends on the minimum temperature achieved within the tumor. Thermal distribution in tumors depends on the vascularity and the extent of hypoxic core besides heating technology. Those tumors which can be heated well have a chance to respond better while those which remain cold will have an inferior outcome.
Current treatment planning systems for hyperthermia are based on computed tomography (CT) perfusion maps and dielectric constants and are still far from being perfect. Texture-based radiomics analysis of CT may be a novel approach to select patients who are more likely to respond to treatment.
In this pilot retrospective study, we aim to assess the ability of CT texture-based radiomics analysis to predict response in locally advanced head-and-neck cancer patients treated with chemoradiation and hyperthermia.
| Materials and Methods|| |
Patients with locally advanced head-and-neck cancer who were treated with multiple modalities of chemoradiation and hyperthermia or radiation with hyperthermia with curative intent were included in this pilot retrospective study. Patients treated between January 2018 and December 2019 were included for predicting treatment response using texture-based radiomics analysis of pretreatment CT.
Clinical data of patients were retrieved from the electronic medical archives. The demography of patients, details of treatment, and responses were recorded. All the patients underwent radical radiation with intensity-modulated radiation with 6 MV photons, to a dose of 66–70 Gy in 6–7 weeks. Hyperthermia was delivered using modified Thermotron radiofrequency, Vinita, Japan. Patients received weekly cisplatin of 50–60 mg. Acute side effects were documented but not included in the analysis.
Planning CT scans were retrieved from the archives. All patients had undergone CT with intravenous contrast on Siemens Biograph 64 - slice PET/CT scanner.
The initial response was assessed as per the RECIST 1.1 criteria. The response was generally assessed within a week of the conclusion of the treatment by comparing gross tumor volume in pretreatment CT scan with that in follow-up CT images. The initial response was scored as complete response (CR), partial response, no response, or progressive disease.
Computed tomography texture analysis
Texture-based radiomics analysis is a technique to objectively quantify tumor heterogeneity on routinely acquired radiological scans (CT, magnetic resonance imaging, etc.) by analyzing the distribution and relationship of pixel or voxel gray levels, not generally perceptible to the naked eye.
For each patient, regions of interest (ROIs) were segmented using manual and/or semi-automated delineation of the largest cross-section slice of the tumor (central core would represent the hypoxic component of the tumor, and hence, important for assessing the response) well as top and bottom slices (2 cm from either pole) representative tumor sections. Each ROI underwent the computed tomography texture analysis (CTTA), which comprised a filtration-histogram-based technique where the filtration step extracted and enhanced features of different sizes corresponding to the spatial scale filter (SSF) which varied from fine (SSF = 2 mm), medium (SSF = 3–5 mm), and coarse (SSF = 6 mm) texture scales. Quantification of texture postfiltration was undertaken using statistical and histogram metrics, namely mean intensity (average brightness), standard deviation ([SD] dispersion from average intensity), entropy (reflects irregularity), mean of positive pixels (average brightness considering only pixels with positive values), skewness (reflects asymmetry of the histogram distribution), and kurtosis (reflects sharpness or pointedness of the histogram distribution). Quantification of texture without filtration on the conventional CT image was also employed as a control. ROI segmentation and CTTA described above was undertaken using commercially available research software TexRAD (Feedback Medical Ltd., Cambridge, UK). Computer simulation and phantom study describe what does the filtration-histogram-based CTTA actually means and how it reflects different components of tissue/tumor heterogeneity in terms of an object/feature size, number/concentration, and variation in the intensity of the objects/features in relation to the background tissue/tumor.
A number of studies using filtration-histogram-based CTTA have been undertaken qualifying it as an imaging biomarker in terms of biological correlates, technical validation, clinical validation (prognosis, disease severity, and treatment response/prediction in a number of oncological applications), clinical adoption, utility, and cost-effectiveness.
The ability of the radiomics parameters (CT size, CT density, and CTTA) to differentiate responders from nonresponders (NRs) were evaluated using the nonparametric Mann–Whitney test. A composite score was developed, combining the significant univariate markers, and its ability to predict responders from NR was assessed and compared to the individual markers. The receiver operating characteristics curve established the diagnostic criteria (area under the curve [AUC], sensitivity, specificity, and P value). Statistical analyses were performed using SPSS (IBM Corp. Released in 2017. IBM SPSS Statistics for Macintosh, version 25.0. IBM Corp., Armonk, NY, USA). P < 0.05 was considered statistically significant.
| Results|| |
A total of 25 patients' data were analyzed of which 15 were males. The average age of the patients was 63.32 years.
[Table 1] shows the anatomical subsites of involvement. Patients were staged according to the TNM (AJCC staging system of classification). Most of the patients had either Staged IV or III diseases [Figure 1]. Only four patients received ≤62 Gy of radiation [Table 2]. Hyperthermia was delivered weekly. Six of the patients received <5 sittings; eight patients received 5–10 sittings, and 11 patients received ≥10 sittings but <14 sittings [Table 3] of hyperthermia. CR was seen in 16 patients, six had a partial response, and one did not respond while two patients progressed [Table 4]. They were further categorized as responders (CR) and NR. Only those who had CR, that is, 16 patients were included as responders and the rest as NR (n = 9).
|Table 3 : Association between hyperthermia cycles and number of patients|
Click here to view
Among the different texture parameters, kurtosis of the lesion (center core) at fine texture scale was a significant differentiator of NR from CR [P = 0.009; [Figure 2] with a higher kurtosis value associated with NR. Particularly, a kurtosis value ≥1.29 identified NR with a sensitivity of 82% and specificity of 69% [AUC = 0.8; P = 0.01; [Figure 3]. Size of the lesion (center core) measured as the total number of pixels on CT was a significant differentiator of NR from CR [P = 0.039; [Figure 4] with a large lesion size associated with NR. Particularly, a size ≥298 identified NR with a sensitivity of 82% and specificity of 62.5% [AUC = 0.7; P = 0.038; [Figure 5]. The mean density of the lesion (top section) of CT (Hounsfield unit [HU]) was a significant differentiator of NR from CR [P = 0.044; [Figure 6], with a higher HU associated with NR. Particularly, a mean density HU ≥47.4 identified NR, with a sensitivity of 82% and specificity of 75% [AUC = 0.7; P = 0.043; [Figure 7].
|Figure 2: kurtosis at fine texture scale for the center core is a significant differentiator between complete response and nonresponders. A higher kurtosis value is associated with no response|
Click here to view
|Figure 3: Receiver operating characteristics analysis further demonstrated kurtosis at fine texture scale ≥1.29 could identify nonresponder with 82% sensitivity and 69% specificity with a area under the receiver operating characteristics curve (AUC) of 0.8, P = 0.01|
Click here to view
|Figure 4: Size (total number of pixels) of the center core is a significant differentiator between complete response (n = 16) and nonresponder (n = 11) (P= 0.039 – nonparametric Mann–Whitney test. A larger size is associated with nonresponder|
Click here to view
|Figure 5: Receiver operating characteristics analysis further demonstrated size ≥298 could identify nonresponder with 82% sensitivity and 62.5% specificity with a area under the receiver operating characteristics curve (AUC) of 0.7, P = 0.038|
Click here to view
|Figure 6: Mean density Hounsfield unit of the top section is a significant differentiator between complete response (n = 16) and nonresponder (n = 11) (P= 0.044 – nonparametric Mann–Whitney test). A higher mean density value is associated with nonresponder|
Click here to view
|Figure 7: Receiver operating characteristics analysis further demonstrated mean density Hounsfield unit ≥47.4 could identify nonresponder with 82% sensitivity and 75% specificity with a area under the receiver operating characteristics curve (AUC) of 0.7, P = 0.043|
Click here to view
A novel composite score combining the significant univariate markers – CT fine texture quantified as kurtosis, CT size, and CT density (HU) was the best differentiator of NR from CR [P < 0.001; [Figure 8] with a higher score associated with NR. Particularly, a composite score ≥2 identified NR with a sensitivity of 82% and specificity of 81% [AUC = 0.9; P = 0.001; [Figure 9].
|Figure 8: Composite score is the best significant differentiator between complete response (n = 16) and nonresponder (n = 11) (P< 0.001 – nonparametric Mann–Whitney test). Box-plot: complete response = 0 and nonresponder = 1. A higher composite score is associated with nonresponder|
Click here to view
|Figure 9: Receiver operating characteristics analysis further demonstrated composite score ≥2 could identify nonresponder with 82% sensitivity and 81% specificity with a area under the receiver operating characteristics curve (AUC) of 0.9, P = 0.001|
Click here to view
| Discussion|| |
The scope of imaging has transcended the traditional morphological description of an image. Functional imaging, spectroscopy, elastography, and diffusion-weighed imaging offer information beyond morphological features. Texture-based radiomics analysis is another modality of image analysis, which has gained attention in recent years. Texture analysis can provide quantitative information related to tumor heterogeneity by analyzing the distribution and relationships of pixel/voxel gray scales in images. There are four principle methods of texture analysis. They are statistical, structural, transform, and model based. Each of the methods varies in complexity and the need for large computational capabilities.
The statistical method is simple and more popular besides being validated. The method generally referred to as first-order statistics, reflects the gray-level frequency distribution from a histogram involving every pixel intensity of tumor. The parameters used to assess pixel intensities are SD, skewness; kurtosis studies a run-through matrix which can objectify texture heterogeneity of an image in a specific direction. The gray-level co-occurrence matrix is another technique of the second-order statistics, which describes how often a pixel with a specific value pair in a specified spatial range of an image.
Predicative assays to assess the response following radiation with hyperthermia have received less attention. Pretherapeutic classification of tumors as potential responders and NR can assist in choosing the right patients for the addition of hyperthermia to radiation or chemotherapy. There are multiple variables such as size, vascularity, size of the hypoxic core, and dielectric constant of tissues, which can influence the outcome. It is proposed that texture-based radiomics analysis can be additional noninvasive methods to classify potential responders from NR.
Urlich et al. assessed the potential of radiomic features obtained at base live from 18F-Fluorothymidine Positron Emission Tomography (PET) in patients of head-and-neck cancer. They concluded that the lesions which were more homogenous at baseline had a better prognosis.
Bogowicz et al. studied PET-CT in 128 head-and-neck cancer patients. They concluded that most homogenous tumors in CT density with a focused region of high fluorodeoxyglucose intake indicated a better prognosis. CT-based texture analysis, however, overestimated the probability of tumor control in those patients deemed with a poor prognosis group.
Kumo et al. assessed radiomic features in 62 patients of head-and-neck cancer. The assessment was based on CT scan, as in this study. They looked at the multivariate analysis of three histogram features such as geometric mean, harmonic mean, fourth moment and four gray-level-run length features, and short-run emphasis were significant predictors of outcome after adjusting for clinical variables.
Smith et al. have shown the ability of filtration-histogram-based texture analysis (as employed in this study) to be an independent predictor of overall survival in patients with locally advanced squamous cell carcinoma of the head and neck who were treated with induction TPF chemotherapy.
This is the first study discussing the possibility of texture analysis to identify responders from NR in head-and-neck cancer patients treated with hyperthermia in addition to radiation or chemotherapy. The composite score with kurtosis at fine texture scale, size, and density on CT has shown potential in identifying responders from NR before starting the treatment. This small pilot study demonstrates the possibility of texture-based radiomics analysis as a predictive assay in patients receiving hyperthermia with radiation and chemoradiation. A future study should validate these preliminary findings in a prospective study comprising a larger patient population.
| Conclusion|| |
Texture-based radiomics analysis on CT is an inexpensive, noninvasive addition based on imaging for prognostication potentially acting as an imaging biomarker. A composite score – a new paradigm as illustrated in this article demonstrates promise in assessing response in head-and-neck cancer patients treated with hyperthermia in addition to radiation or chemotherapy, which needs to be further confirmed in a larger, prospective study.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Nagraj G Huilgol, Gopal Pemmaraju, Rohit Dusane. Promising role of hyperthermia concurrent with radiation in the treatment of head and neck cancers: A retrospective comparitive study.J Radiat Cancer Res 2019;10:165-9.
Ganeshan B, Miles KA, Quantifying tumours hetrogenitey with CT. Cancer imaging. 2013:13:140-9.
kennets A. Miles, Ganeshan B, Hayball MP. CT texture analysis using the filtration histogram method. What do the measeurements mean? Cancer Imaging.2013;13:400-6.
Lubner MG, Smith AD, Sandrasegaran K, Sahani DV, Pickhardt PJ. CT texture analysis: definitions, applications, biologic correlates and challenges. Radiographics. 2017;37:1483-503.
Urlich EJ, Menda Y, Ponto LL, Anderson CM, Smith BJ, Sunderland JJ, et al
. FLT-PET- Radiomics for response predication to chemoradiaiton therapy in Head and Neck squamos cell cancer. Tomography 2019;5:161-9.
BogomiCZ M, Reisterer O, Stark L S, Studer G, Unkelbach S, Gukenenberger M, Thandini- Lang S. Comparison of PET and CT radiomics for prediction of local control in head and neck squamous cell carcinoma. Acta oncOL. 2017;56:1531–6.
Kumo H, Qureshi M M, Chapman M N, B Liv, V.C. Andren-Arsa, k Onoue K, M.T. Truong, Sakai O.CT texture Analysis potentially predicts local failure in Head and Neck Cancer squamous cell carcinoma treated with chemo radiotherapy.AJNR. Am. J. Neuroradiology: 2017;38:2334-40.
Smith AD, Gray MR, del Campo SM, Shlapak D, Ganeshan B, Zhang X, et al.
Predicting overall survival in patients with Metastatic Melanoma on Anti angiogenic Therapy and RECIST stable disease on initial post therapy images using CT Texture Analysis. AJR AM J Roentgenol.2015;205:W283-93.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9]
[Table 1], [Table 2], [Table 3], [Table 4]