Abstract
Objectives: To differentiate squamous cell hyperplasia (SCH) (benign) from squamous cell carcinoma (SCC) malignant) using textural features extracted from CT images and thereby, facilitate the preoperative medical diagnosis and treatment of throat cancers without the need for sample biopsies.
Methods: In total, 100 throat cancer patients were selected for this retrospective study. The cases were collected from the Second Hospital of Jilin University, Changchun, China, from June 2017 to January 2019. The patients were separated into a training and validation cohort consisting of 70 and 30 cases, respectively. The Artificial Intelligence Kit software (A.K. software) was used to extract the radiomics features from the CT images. These features were further processed using the minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) methods to obtain a subset of optimal features. The radiomics model was validated based on area-under-the-curve (AUC) values, accuracy, specificity, and sensitivity using the R-studio software.
Results: The diagnostic accuracy, specificity, PPV, NPV, and AUC values obtained for the training cohort was 0.91, 0.9, 0.93, 0.9, and 0.96 CT angiography (CTA), 0.93, 0.93, 0.95, 0.90, and 0.96 computed tomography normal (CTN), and 0.92, 0.87, 0.91, 0.96, and 0.96 CT venogram (CTV). These values were subsequently confirmed in the validation cohort.
Conclusion: The radiomics-based prediction model proposed in this study successfully differentiated between SCH and SCC throat cancers using CT imaging, thereby facilitating the development of accurate preoperative diagnosis based on specific biomarkers and cancer phenotypes.
Head and neck squamous cell carcinomas (HNSCC) represent the fifth most common cancers in the world as compared to other types of cancers.1,2 The HNSCC group of cancers includes oral, nasopharyngeal, and throat cancers,3 in which throat cancer is the second most common cancer affecting the respiratory tract with a survival rate of 64.2%. In general, cancer therapy includes surgery, radiotherapy, chemotherapy, or biological agents that have been selected according to the patient’s case. The risk factors for throat cancer infection increases with alcohol consumption and cigarette smoking, in addition to other factors such as genetics and nutrition.4 For throat cancer diagnosis, a CT scan is usually recommended by the radiologist as it is the fastest method to obtain a result. The use of thin-slice helical CT with 2-dimensional (2D) reconstruction is commonly employed for the diagnosis. During this scanning process, the appearance of soft tissue and bony details such as calcification, abnormal changes in the throat, and site of the spinal canal can be clearly demonstrated.5 The differentiation between cystic or necrotic tissues is supported by CT angiography (CTA) and CT venogram (CTV) scans, whereby the tumor regions appear brighter due to their heterogeneous enhancement pattern as opposed to the normal regions.6 Nevertheless, the radiologist faces a challenge in the diagnosis of throat cancers especially in the early-stages due to non-specific symptoms such as hemorrhage, calcification, and presence of heterogeneous masses. The tumor histology is usually determined using biopsy, although this method may produce errors in sampling or region detection and subsequently result in complications.7 Therefore, using CT imaging, selected features from the clinical images of the tumor can be extracted to discriminate intra-tumor heterogeneity. The differentiation between the intratumor heterogeneity has a significant effect on the prediction, diagnosis, prognosis, response, and staging of cancer.8 Radiomics is a powerful technique employed to analyze image textures. It is a new method in translational research that has been applied for the discrimination of cancer phenotypes using quantitative features derived from the imaging data. For instance, a high number of textural features such as measurements from the mean and standard deviation can be extracted from the imaging data and converted into predictive mathematical models9 that may facilitate clinical decisions. Radiomics is a useful method to differentiate between benign and malignant lesions in head and neck cancers, and early metastatic lesions as well as provide prognosis and predictive factors for the tumor, thereby leading to an effective treatment plan for patients.
Usually, throat cancers are discovered in the later stages since it is difficult to diagnose this cancer in the early stages, thus decreasing the rate of survival. Therefore, radiomics is applied to the medical imaging analyses on existing tumors or newly diagnosed tumors of patients as it may provide new information to physicians. For instance, the expression of genomic and proteinuria systems based on the features of the radiomics model for these images can be extracted; thus, allowing physicians to obtain phenotypic gene-protein patterns which may include prognostic information.10,11 Therefore, this additional information is beneficial and provides details that may not be visible to the naked eye. This study aims to differentiate squamous cell hyperplasia (SCH) (benign) from squamous cell carcinoma (SCC) (malignant) using textural features extracted from CT images and thereby, facilitate the preoperative medical diagnosis and treatment of throat cancers without the need for sample biopsies.
Methods
This retrospective throat cancer diagnostic study was approved by the Second Hospital Ethical Committee of Jilin University, Changchun, China. The informed consent was waived off because of the nature of the study, which was retrospective, with preserving patients’ data and not disclosing and handled with strictly confidential. This study has been conducting based on the Helsinki Declaration. The medical imaging technology used in this study was based on the picture archiving and communication system (PACS) obtained from the Second Jilin University Hospital, Changchun, China. The images were reviewed from June 2017 to January 2019 to select patients who were pathologically diagnosed with throat cancer. This study focused on patients that underwent CT scanning, in which the medical history of patients such as gender, smoking, alcohol, tumor staging, and histopathology of the surgical specimen was available in the database. In total, 100 patients with throat cancer were selected for the training and validation groups, in which 60 patients were diagnosed with SCC and 40 patients were diagnosed with SCH. Based on their gender and age group classification, 73 males and 27 females, ranging from 44 to 85 years, were selected as participants and the Kolmogorov-Smirnov test was applied to determine the normal distribution of the data. Information regarding the lifestyle choices of the patients was also recorded and the patient profile consisted of smokers (n=74), non-smokers (n=14), and those who quit smoking (n=12). In addition, the participants also comprised alcoholic-drinkers (n=73), non-drinkers (n=20) and those who stopped drinking (n=7). This study established exclusion and inclusion criteria and followed with high strictness and transparency to know these criteria. Figure 1 depicts the classification of the study participants and clarifies the mechanism followed to exclude and include patients.
- Flowchart representing the patient selection criteria process. SCH: squamous cell hyperplasia, SCC: squamous cell carcinoma (SCC)
Computerized tomography (CT) scan parameters
The CT scans of the throat were performed using a 256-Slice scanner (Philips iCT, Netherland), with the tube voltage and tube current set at 120 kVp and 200 mAs, respectively. The scan parameters were established as follows: scanning layer thickness of 1 mm in diameter, scanning layer spacing of 1 mm, pitch ratio of 0.342, matrix array dimension of 256 x 256, and scan time of 1-3 seconds. A preoperative throat scan was first obtained using CTN imaging that was performed without contrast media. The second CT scan was accomplished for CTA 35s and CTV 65s using contrast media as well as a non-ionic iodinated contrast medium for intravenous injection (dosage concentration of 1.5 mL/kg and injection rate of 3.5 mL/s) (Ultravist, 100 Bayer Blvd, Whippany, NJ, United States). The reconstruction was performed for all the throat scans according to the established criteria. The CT images were exported to DICOM format for feature extraction.
Setup experiments
The CT images of SCH and SCC were segmented manually based on the region of interest (ROI) by a radiologist with 3 years of experience. Subsequently, the radiomics features were extracted from these segmented CT images. The ROI was reviewed by an expert radiologist with more than 20 years of experience. The ROI was drawn around the tumor outline walls for all the images. The borders of the tumor were excluded from this study as the edges of the tumor can sometimes merge with normal tissues and subsequently produce weak results. The ROI’s artificial outline for the CT images was performed using ITK-SNAP 3.8.0 (www.itksnap.org)12,13 (Figure 2).
- The segmentation process by ITK-SNAP software for squamous cell carcinoma showing A-C) malignant and D-G) the squamous cell hyperplasia benign.
The radiomics features obtained using the A.K. software (GE Healthcare, China) were automatically calculated and divided into volumetric, texture, histogram, and wavelet features.14 The radiomics data provided 396 features, although most of them were not useful for clinical applications. For the selection of useful features with potential clinical application, least absolute shrinkage and selection operator (LASSO) and minimum redundancy maximum relevance (mRMR) methods were employed to eliminate redundant and irrelevant features before proceeding with the construction of the logistic regression model. The mRMR method was used first, and followed by the LASSO method which was used as a confirmation method. These features were associated with one or more useful outcomes under the logistic regression model. These features were subsequently divided into 6 groups consisting of a histogram, gray level size zone matrix, form factor matrix, gray-level co-occurrence matrices, and run-length matrix for data analysis.
For the selection process using the mRMR method, 30 features were retained in this process and the irrelevant features were eliminated. Next, the LASSO method was performed to extract the most effective subset of features for the construction of the final logistic regression model. In the LASSO model, a 10-fold cross-validation analysis was used to create the optimal penalization coefficient lambda (λ). The λ value that gave the lowest average binomial deviance was used in the feature selection. The vertical lines accompanied by red points were drawn for the optimal values using the lowest criteria. Lastly, a coefficient profile plot was produced against the log (λ) sequence based on the coefficient LASSO profiles of 1188 features.
Statistical analysis. The statistical analyses for the extracted features obtained from the CTA, computed tomography normal (CTN), and CTV images were performed using the open-source R-Studio Server software. The values for the multivariate logistic model were considered significant at p<0.05. The receiver-operating-characteristic (ROC) and area-under-the-curve (AUC) analyses were used to evaluate these models. The logistic regression model and ROC curve were analyzed using the R software package to discriminate SCH from SCC lesions from the training cohort and validation cohort. The optimal threshold value calculated for the training cohort was applied to the validation cohort to differentiate the SCH lesions from SCC lesions. Therefore, the optimal threshold was obtained by maximizing the Youden index. All the statistical analyses were performed using R-Studio Server (Version 1.1.463, GE Healthcare, China).
Results
The data in this study was obtained from 100 patients who suffered from throat cancer. Based on their histology and pathology results, 60 patients suffered from SCC and the remaining 40 patients suffered from SCH. These patients were divided into the training group (70%) and test group (30%). Additionally, no statistical differences were observed between the age (p=0.12), gender (p=0.50), smoker status (p=0.90), and alcohol consumption (p=0.90) for both groups.
Construction of a radiomics model and feature extraction from CT images
In total, 1188 features were extracted from the CTA, CTN, and CTV images, whereby 396 features were extracted from each segmented ROI image. Once the number of features were stabilized, the most effective subset of the extracted features was determined to evaluate the corresponding coefficients. A total of 12 features were selected from the extracted features for CTA, 11 for CTN, and 12 for CTV; thereby, indicating the construction of a promising radiomics-based prediction model for throat cancer.
The diagnostic capability of the CT radiomics model for throat cancer
In this study, CTA, CTN, and CTV parameters were used for the CT images to measure the diagnostic abilities of these parameters based on the receiver operating characteristic (ROC) curve values. The dataset obtained using these parameters were trained separately using the logistic regression model. The radiomics model constructed based on the CT images resulted in the AUC value of 0.96 (95% CI: 0.92-1) for all 3 parameters (CTA, CTN, and CTV) in the training cohort.
The logistic regression model was then tested using the validation cohort for all 3 parameters. The radiomics model based on the CT images resulted in the AUC value of 0.94 (95% CI: 0.82-1) for CTA and CTV, while a higher AUC value of 0.96 (95% CI: 0.87-1) was obtained for the CTN parameter.
The predictive performance of the different CT images and comparison of radiomics features extracted from these images such as accuracy, specificity, sensitivity, positive predictive value (PPV), and negative predictive values (NPV) for the training cohort and validation cohort are shown in Table 1.
- The performance of the CT parameters based on the radiomics model.
Based on the results obtained, the CTN radiomics model provides a promising diagnostic ability for throat cancer as compared to the other parameters of the CT images. Specifically, the logistic regression model produced significantly higher results for AUC (0.96), accuracy (0.93), sensitivity (0.93), and specificity (0.93) in the training cohort, and these results were subsequently confirmed in the validation cohort with values of 0.96, 0.84, 0.94, and 0.33, respectively. However, the CTN radiomics model yielded better results due to the absence of contrast media, while the contrast enhancement displayed in CTA and CTV introduced heterogeneity in the radiomics features and decreased their predictive ability.
The mean AUC values for all 3 parameters used in this study are represented in Figure 3. The mean AUC value for CTN was higher compared to the values obtained for CTA and CTV imaging. A statistically significant difference was also observed between the values for all 3 parameters (p<0.0001).
- Box plot of area-under-the-curve (AUC) values for throat cancer illustrating the higher values of computed tomography normal (CTN) parameters as compared to computed tomography artery (CTA) and computed tomography venous (CTV).
The predictive performance of the radiomics model proposed in this study was higher as compared to the clinical model. Figure 3 depicts the nomogram for the decision curve based on the evaluation of the clinical application of the imaging data. The net benefit in the y-axis of the nomogram is dependent on the difference in the expected benefit and the expected harm for different threshold probabilities in the validation cohort based on the introduced model [Net benefit = true positive rate - (false positive rate × weighting factor), weighting factor = threshold probability/(1-threshold probability)].
The analysis of extracted features
Based on the feature selection by the mRMR and LASSO methods, only the most significant features identified in the logistic regression model for each parameter were used in this study. In total, 12 features (Inverse Difference Moment, Cluster Shade, GLCM Energy, Haralick Correlation, Volume, Short Run Emphasis, Long Run High Grey Level Emphasis, Correlation, Haralick Correlation) were selected for CTA (Figure 4A), while 11 features (Cluster Prominence, Cluster Shade, GLCM Energy, Haralick Correlatio, Correlation, Short Run Low Grey Level Emphasis, Inverse Difference Moment), and 12 features (Cluster Shade, GLCM Energy, Correlation, Inertia, Surface Area, Run Length Nonuniformity, GLCM Entropy) were selected for CTN (Figures 4B) and CTV (Figure 4C), respectively. The features selected using the LASSO model parameters varied significantly between SCH lesion and SCC. The statistically significant values for the training cohort and validation cohort are listed in Table 2.
- The least absolute shrinkage and selection operator (LASSO) features after univariate logistic regression and removing redundancies, for each parameter was used in our study. A) The right and left image of compared to computed tomography artery (CTA), B) the right and left image of computed tomography normal (CTN), C) the right and left image of computed tomography venous (CTV).
- The statistically significant for least absolute shrinkage and selection operator (LASSO) features selected of throat cancer for each parameter used in this study.
Discussion
In this study, 3 parameters (CTA, CTN, and CTV) were used for the CT images based on the logistic regression model as compared to a previous study that used only one parameter.15 The results obtained in this study yielded statistically significant differences between the SCH and SCC lesions based on the 3 parameters used as compared to previous studies.16,17 In these studies, the authors proposed the use of computer-aided diagnosis (CAD) for the prediction of patients with throat cancer, thereby increasing the diagnostic performance for early-stage cancers18 as well as minimizing errors using a classification system. In this study, values of 0.78 were obtained for CTA, 0.84 for CTN, and 0.78 for CTV. Data from the CTN imaging resulted in a higher accuracy, specificity, and sensitivity in the training cohort, in which the data was subsequently confirmed in the validation cohort based on the established parameters. Thus, these results indicate that the diagnostic evaluation established in this study was relatively stable. Additionally, the discrimination of SCH lesions from SCC throat cancers using CT images was supported by the development of a radiomics score. The radiomics score consisted of a nomogram that performed well in differentiating between benign lesions and malignancy. The prediction based on the calibration curve for throat cancer and the decision curve analysis (DCA), which included a nomogram, were shown to be useful for clinical application.19
Despite the importance of CTA and CTV in clinical cases, their predictive ability in the radiomics model was lower due to the use of contrast enhancement that resulted in heterogeneity for the radiomics features. No enhancement is required for CTN images, thus resulting in its higher predictive ability as compared to other parameters when used in the radiomics model.
Radiomics is an emerging field of research based on computer-aided technology, in which medical images are converted into a sequence of data using computer algorithms.20,21 The microscopic properties of the tumor biomarkers and biological phenotypes are captured by the image and the features of the tumor are extracted using radiomics software to identify the heterogeneity of the tumor biomarkers.22,23 The relationship between the extracted features and biological phenotype is quite complex,24 mainly due to the extraction of hundreds of radiomics features for these tumor biomarkers. Therefore, the utility of these features is evaluated based on the parameters established for the data.25 The radiomics score was developed in this study to determine the prognosis of SCH lesions from SCC lesions. The differentiation, calibration, and clinical utility of this score was applied using a training cohort and subsequently, a validation cohort to evaluate the performance of this model. The radiomics model achieved a good result as compared to the conventional model.
The radiomics features derived from the CT images were quantitatively evaluated in this study to discriminate SCH lesions from SCC throat cancer lesions. Additionally, a radiomics-based prediction model was developed using the logistic regression model. The CTN scan significantly discriminated SCH lesions from SCC throat cancers, as indicated by the AUC values of 0.96 obtained in the training cohort and 0.96 in the validation cohort. The results observed in this study were consistent with previous studies indicating that normal CT images exhibited a higher prediction performance as compared to enhanced CT images.26-28 The main reason for this observation is attributed to the clearer identification of the phenotypic heterogeneity of tumors using the normal CT scan as opposed to the reduced visibility of enhanced CT images due to the existing intra-tumoral contrast material. Specifically, the ROI contains the entire lesion, as opposed to previous studies that have only ≥1 cm.29 The patterns and distributions such as cystic and calcification 30 are crucial for the differentiation between benign and malignant tumors.31
This study also offers an advantage that has not been previously described in any study. Two methods were used to select the features required for the 3 CT images, as opposed to the use of only a single method LASSO.18,29 For instance, the mRMR method was performed to eliminate redundant and irrelevant features. Subsequently, the LASSO method was used to select the optimized subset.32 The features were selected with high accuracy and were able to significantly discriminate SCH from SCC lesions, as evidenced by the results in this study.
Study limitations
The number of sample cases included in this retrospective study was insufficient as only 100 cases of SCC and SCH were analyzed. Although the Digital Imaging and Communications in Medicine (DICOM) data obtained by the single-center, single CT can ensure data consistency, the ability to performing a comprehensive evaluation of the model is lacking. Finally, the manual sketching of ROI by radiologists can lead to discrepancies. Therefore, it is envisaged that future studies should incorporate a larger sample size using multi-center radiomics data.
In conclusion, the radiomics features offer a rapid topography and biomarker assessment of cancer. Using this approach, the use of medical imaging, computer software, and machine learning are combined to obtain a clearer vision of cancer. The improvement of medical imaging data, together with the development of a new computational model, support the diagnosis and treatment of cancers with high accuracy. Cancer research is essential for the discovery of tumor heterogeneity. Hence, it is necessary to develop a computational model that optimizes the significant radiomics features for its application in clinical cases.
This study has yielded promising results for the differentiation of SCH lesions from SCC throat cancers based on the 3 parameters established for CT imaging. Moreover, the CTN data imaging was shown to have more efficiency as compared to the other parameters identified for CTA and CTV. It is envisaged that the findings of this study can facilitate the early diagnosis of throat cancer without the need for biopsies, thereby leading to higher survival rates and faster administration of medical treatments.
Acknowledgment
This work was supported by the Technology Development Plan for Jilin Province, China (No. 2018010119JC), the Youth Training Foundation by the Health and Family Planning Commission of Jilin Province, China (No. 2017Q012), and Health Special Project of Jilin Province Department of Finance, China (No. 3D518V283429). We would like to acknowledge Proofreading by A UK PhD company (www.proofreadingbyphd.com) for a marvelous English language editing service.
Footnotes
Disclosure. Authors have no conflict of interests, and the work was not supported or funded by any drug company. This study was supported by the Technology Development Plan for Jilin Province, China (No. 2018010119JC), the Youth Training Foundation by the Health and Family Planning Commission, Jilin Province, China (No. 2017Q012), and the Health Special Project of Jilin Province Department of Finance, Jilin Province, China (No. 3D518V283429).
- Received October 2, 2020.
- Accepted February 1, 2021.
- Copyright: © Saudi Medical Journal
This is an open-access article distributed under the terms of the Creative Commons Attribution-Noncommercial License (CC BY-NC), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.