A Deep Learning Radiomics Analysis for Survival Prediction in Esophageal Cancer

Wang, Junxiu; Zeng, Jianchao; Li, Hongwei; Yu, Xiaoqing

doi:https://doi.org/10.1155/2022/4034404

Journal of Healthcare Engineering

On this page

Abstract Introduction Materials and Methods Results Discussion Conclusions Data Availability Conflicts of Interest Acknowledgments Supplementary Materials References Copyright Related Articles

Special Issue

Explainable Artificial Intelligence for Medical Applications

View this Special Issue

Research Article | Open Access

Volume 2022 | Article ID 4034404 | https://doi.org/10.1155/2022/4034404

A Deep Learning Radiomics Analysis for Survival Prediction in Esophageal Cancer

Junxiu Wang,^1,2Jianchao Zeng ,¹Hongwei Li,³and Xiaoqing Yu¹

Academic Editor: Enas Abdulhay

Received11 Jan 2022

Accepted03 Mar 2022

Published24 Mar 2022

Abstract

The purpose of this study was to explore the deep learning radiomics (DLR) nomogram to predict the overall 3-year survival after chemoradiotherapy in patients with esophageal cancer. The 154 patients’ data were used in this study, which was randomly split into training (116) and validation (38) data. Deep learning and handcrafted features were obtained via the preprocessing diagnostic computed tomography images. The selected features were used to construct radiomics signatures through the least absolute shrinkage and selection operator (LASSO) regression, maximizing relevance while minimizing redundancy. The DLR signature, handcrafted features’ radiomics (HCR) signature, and clinical factors were incorporated to develop a DLR nomogram. The DLR nomogram was evaluated in terms of discrimination and calibration with comparison to the HCR signature-based radiomics model. The experimental results showed the outperforming discrimination ability of the proposed DLR over the HCR model in terms of Harrel’s concordance index, 0.76 and 0.784, for training and validation sets, respectively. Also, the proposed DLR nomogram calibrates and classifies better than the HCR model in terms of AUC, 0.984 (vs. 0.797) and 0.942 (vs. 0.665) for training and validation sets, respectively. Furthermore, the nomogram-predicted Kaplan–Meier survival (KMS) curves differed significantly from the nonsurvival groups in the log-rank test ( value <0.05). The proposed DLR model based on conventional CT images showed the outperforming performance over the HCR signature model in noninvasively individualized prediction of the 3-year survival rate in esophageal cancer patients. The proposed model can potentially provide prognostic information that guides and helps the clinical decisions between observation and treatment.

1. Introduction

Esophageal cancer (EC) is the eighth most common malignancy and the sixth most common disease-related cause of death worldwide [1, 2]. The incidence of esophageal cancer is notably high in Asia and Iceland, as well as the United Kingdom and the United States [3–5]. In order to systemically control the disease, radiotherapy and neoadjuvant chemotherapy are commonly combined with surgery [6–8]. Regardless of its benefit for severe patients with a low survival rate, an aggressive treatment plan, including multiple cycles of treatment and adjuvant chemotherapy, is not suitable for the other patients with esophageal cancer [9]. Pre-identification for such patients having a low survival rate before surgery can help provide other suitable treatment regimens for these patients [10]. Therefore, identification of patients having a lower survival rate is vital to take benefit from additional treatment.

Radiomics features [11, 12] have been widely used as an extremely useful tool in quantitative analysis of medical imaging and in medical diagnosis [13, 14]. The traditional radiomics utilized handcrafted features, such as tumor shape and texture, obtained from medical images [11]. However, such handcrafted low-order features are not suitable to define intrinsic characteristics of intratumor imaging heterogeneity, limiting the applicability of the radiomics model [11–13]. Furthermore, the construction of handcrafted features is limited within the known knowledge of medical imaging.

Deep learning, especially CNN, has recently achieved promising results in medical image analysis [15–18]. Deep networks consist of multiple layers that can be learned from data [19, 20]. For example, prognostic tumor features can be extracted through hierarchical convolution operation on the medical images [21–24]. Compared with handcrafted features, DLR features contain more important tumor information that may help diagnose [25–27].

However, unlike handcrafted radiomics studied widely for radiological diagnosis and prediction [28], the application of deep learning in predicting overall survival in esophageal cancer has not been thoroughly explored yet. Hence, this study aims to develop and validate the deep survival prediction model based on a radiomics nomogram for individualized prediction of three years’ overall survival in patients with esophageal cancer.

It is worthwhile to highlight three aspects of the contributions here. (1) This study investigated DLR features in the survival prediction of patients with esophageal cancer. Unlike the traditional handcrafted features, clinical target-oriented DLR features can be automatically learned from data. (2) Both HCR features and DLR features are considered in the prediction model to characterize the esophageal lesions thoroughly. The main reason is that the DLR and HCR can describe imaging heterogeneity of esophageal lesions in different levels. In particular, HCR comprise shape features, first-order statistics, and texture features; DLR contains the “real-world” textures, which are extracted from the pretrained DenseNet-169 network via transfer learning strategy. (3) This study develops a noninvasive predictive model that combines deep learning-based radiomics features, handcrafted features, and clinical factors to predict survival rates within three years at diagnosis of esophageal cancer. The DLR nomogram survival prediction of esophageal cancer patients can allow more proper treatment. The experimental results showed that the DLR nomogram outperforms the HCR model and the clinical model.

2. Materials and Methods

2.1. Patients

Esophageal cancer patients at Shanxi Cancer Hospital were the subject of our retrospective study. The patients were included according to the inclusion criteria: (a) patients who had pathologically confirmed esophageal cancer, (b) a standard CT scan performed before any treatment, and (c) clinical characteristics available. The patients were excluded with the following criteria: (a) too poor CT image quality, which may affect the diagnosis of the patient, (b) patients who had chemotherapy treatment at another institution, and (c) patients who are also suffering from other cancers.

The survival group includes patients who survived more than three years since the treatment, whereas the nonsurvival group includes patients who died within three years. A total of 154 esophageal cancer patients diagnosed from November 2012 to February 2015 participated in our retrospective study. Those data were grouped into two sets: training (116) and validation (38) data at a ratio of 3 : 1.

Baseline clinical data were collected via the electronic medical record system (EMRS) [29], including gender, BMI, age, M-stage, N-stage, T-stage, overall stage, and planning target volume (PTV). The picture archiving and communication system (PACS) was used to obtain CT images. The dataset was constructed and evaluated in April 2019, and all enrolled patients were followed for at least 3 years. The Institutional Review Board approved the study.

2.2. CT Image and Region of Interest (ROI) Acquisition

General Electric Light Speed RT16 was used for scanning, with a CT thickness of 5 mm. The primary tumor volumes for radiotherapy planning were set as the ROI to quantitatively analyze the images. Two skilled radiologists manually selected the three-dimensional tumor ROI using the software package 3D Slicer [30].

Training CT images were preprocessed to avoid accuracy degradation of DL models caused by noises introduced with the interval change, which include resampling, rescaling, and voxel normalization. Those CT images were reconstructed with a matrix of 512 × 512 and 0.5 × 0.5 mm² pixel size, and the resampling with cubic interpolation to 1 × 1 × 1 mm³ pixels was conducted, minimizing CT images variabilities [31].

The tumor area was located with a rectangle bounding box that covers the primary tumor area. The ROI for each patient was obtained with three cropped consecutive slices to avoid the bias of manual segmentation that affects the location of a bounding box. Lastly, the tumor image was resized to 224 × 224 × 3 voxels.

2.3. Radiomics Feature Extraction

Phenotypic differences between tumors can be captured by a large number of quantitative radiomics features. In this study, deep learning features and handcrafted features were extracted to quantify tumor phenotype to enhance the learning efficiency of the radiomics model. Those two feature sets have complementary advantages that can be combined to improve the model. Also, expert knowledge on the esophageal cancer lesion can be reflected with shape and texture features. On the contrary, the high-level DLR features can significantly represent complex spatial features in both global and local perspectives.

The handcrafted feature extraction algorithm was standardized by referring to the Image Biomarker Standardization Initiative (IBSI) [32–34] and Radiomics Ontology [35]. For each CT ROI, 1,670 handcrafted features were extracted using Python implementation, including 18 first-order statistics, 16 geometric, and 1,564 texture features. The textural features include 14 gray-level dependence [36], 23 gray-level co-occurrence [37], 16 gray-level run-length [38], 16 gray-level size-zone [39], and 5 neighborhood gray-tone difference [40] matrices. Refer to the supplementary appendix of Lambing [41], for mathematical definitions of those features.

The DenseNet-169, designed for the image classification task, was adopted to extract DLR features. In the training cohort, data augmentation approaches including random rotation, random shear, and random zoom were employed before the training procedure. The deep learning model was pretrained on the ImageNet dataset, one of the largest image datasets, and then fine-tuned in a transfer learning strategy to avoid the overfitting problem [42]. The network was trained with cross-entropy loss function and Adam optimizer with a learning rate of 0.0001, a batch size of 16, and a regularization weight of 0.0001. The network was implemented on Keras (https://keras.io/) with the TensorFlow library as the backend (https://www.tensorflow.org/). As depicted in Figure 1, the tumor ROI was fed into the DenseNet-169, and the outputs of hidden layers were collected to obtain 1,664 features in total.

2.4. DLR Signature Building

In order to obtain the most effective feature, three stages of feature selection were carried out. First, features were obtained through the Mann–Whitney U (MWU) test. Then, the features were sorted based on the mutual information (MI) between features and the survival status using the minimum redundancy maximum correlation (mRMR) scheme [43]. It should be noted that, in this study, only the top 50 features in mRMR were retained. Lastly, the dimension of features is reduced by the LASSO to obtain optimal features [44]. The survival-related features were retained while the other features were removed by LASSO regression. The 10-fold cross validation was conducted with 100 iterations in LASSO regression. The obtained features are used to construct the DLR signature, and the HCR signature was constructed in a similar way for comparison.

2.5. DLR Nomogram Construction

A DLR nomogram was built by integrating DLR signature, HCR signature, and clinical features with a multivariable logistic regression model. Backward stepping selection was used with information criterion of Akaike as the stopping rule [45]. The variable multicollinearity in the multiple logistic regression model was checked by the variance inflation factor (VIF), where VIF > 10 indicates high multicollinearity [46]. A DLR nomogram was then built based on the multivariate logistic analysis, predicting the individual probability of survival in the training dataset.

2.6. Evaluation of the DLR Nomogram

Harrel’s C -index was employed to evaluate the discrimination ability of the DLR nomogram in both training and testing datasets. The bootstrap method was used to resampling 1,000 times, and the C index in both cohorts was calculated with 95% confidence intervals. The AUC, accuracy, specificity, and sensitivity were calculated on the plotted ROC curves. The calibration ability of the DLR nomogram was evaluated using the calibration curve that depicts the consistency between predicted and actual survival probabilities. Hosmer–Lemeshow (HL) test [47] and decision curve analysis (DCA) [48] were utilized to evaluate the fitting accuracy and robustness of the DLR nomogram, respectively. Furthermore, KMS curves were constructed to predict survival status. Accordingly, the patients were predicted as survival or nonsurvival, and then, the difference in survival curves between the two groups was evaluated using the log-rank test.

2.7. Statistical Analysis

All the statistical analyses were conducted with R software (version 4.0.3; http://www.Rproject.org). MWU and Chi-square tests were adopted for univariate analysis, and Spearman’s correlation rank was employed for correlation. The penalty parameter (λ) was tuned by LASSO logistic regression model. This study used the following packages for each analysis. “glomnet” package: LASSO logistic regression, “rms” package: nomograms and calibration plots, “ResourceSelection” package: HL test, “car” package: VIFs calculation, “survivalROC” package: AUC analysis, “survminer” package: KMS analysis, and “dca.R” function: DCA performance. This study utilized a bilateral statistical significance level value <0.05.

3. Results

Figure 2 depicts the schematic diagram of the study.

3.1. Clinical Characteristics

Table 1 summarizes the clinical characteristics of the training and validation cohorts, where the Chi-square test ( value is 0.572) for two data shows no significant observable difference in the survival rate, 30.2% and 36.8%, for training and validation, respectively.

3.2. DLR Signature

Fifty features were obtained for each patient after survival-unrelated and redundant feature removal from 1,664 DLR features. Then, based on the training cohort, 33 potential predictors were selected by LASSO regression. Parameter (λ) selection and coefficients of LASSO are given in Supplementary Material (II). The DLR signature was constructed using the selected features.

The results show that scores of the survival group were higher than the nonsurvival group with a significant difference in terms of DLR signatures (1.10 ± 0.75 vs. −2.22 ± 0.75) in the training cohort and (0.15 ± 1.11 vs. −1.90 ± 1.12) in the validation cohort. MWU test was used with a value <.001. Also, a significant correlation between DLR signature and survival status was found (C index: 0.729, in the training data, and C index: 0.766, all in the validation data).

The LASSO algorithm selected 18 handcrafted features to build HCR signatures. HCR signatures were also significantly different between survival and nonsurvival groups. In the training data, 0.68 ± 1.12 vs. −0.50 ± 1.03, value <.001, MWU test, and in the validation data, 0.80 ± 1.26 vs. −0.23 ± 1.48, value = 0.035, MWU test. HCR feature selection by LASSO regression is described in detail in Supplementary Material (I).

3.3. DLR Nomogram

The DLR signature, HCR signature, and BMI were combined to construct a DLR nomogram, as shown in Figure 3. The VIFs of DLR signature, HCR signature, and BMI were 1.45, 1.41, and 1.07, respectively, indicating no severe collinearity in the regression model.

Figure 4 depicts ROC curves of the DLR nomogram for the DLR signature model and HCR signature model. The AUC was 0.984, 0.955, and 0.797 for the DLR nomogram, DLR signature model, and HCR signatures’ model, respectively, in the training data. In the validation data, the AUC was 0.942, 0.846, and 0.665 for the DLR nomogram, DLR signature model, and HCR signatures’ model, respectively. The results indicate that the DLR nomogram model provides better discrimination ability (Harrel’s concordance index, 0.76 and 0.784, for the training and validation data, respectively).

(a)

(b)

Figure 5 depicts the calibration curves, showing the consistency between predicted and actual survival rates. A nonsignificant statistic of the training cohort ( value = .563, HL test) showed no deviation from the ideal ﬁt. In the validation cohort, the 3-year survival rate was also well-calibrated ( value = .648, HL test).

(a)

(b)

The DCA examined the clinical outcomes based on threshold probability at which a net benefit could be derived. Figure 6 depicts the DCA of the DLR nomogram, showing that the DLR nomogram obtained outstanding net benefits over the other strategies: treat-all-patients and treat-none strategies. A significant difference ( value <.05, log-rank test) between prediction survival and nonsurvival groups was found in KMS curves (Figure 7).

(a)

(b)

(a)

(b)

4. Discussion

Treatment planning can be further individualized via preoperative prediction of three-year survival. In previous studies, handcrafted features were analyzed to predict survival rates. However, due to the limited feature extraction ability, the prediction accuracy was not high enough. In order to overcome such a limitation, this study investigated DLR features in the survival prediction of patients with esophageal cancer. Unlike the traditional handcrafted features, clinical target-oriented DLR features can be automatically learned from data [49].

Intratumor heterogeneity has been considered a potential prognosis factor. The DRL feature extraction can robustly characterize the intratumor heterogeneity noninvasively from the medical images [26]. The experimental results showed that the use of DLR features contributed to the performance of the model, which is also supported by recent studies that high-dimensional features can preserve more detailed cancer information, making them more sensitive when assessing survival status [24]. Therefore, by combining these DLR features, a DLR nomogram survival prediction of esophageal cancer patients can allow more proper treatment. The experimental results showed that the DLR nomogram outperforms the HCR model and the clinical model.

This study has several limitations, described as follows. First, only 154 patients were available for a three-year follow-up analysis. A larger amount of data is required to improve the performance of the model. Second, all the patients were collected to form a single-center, thereby limiting the generalizability of the DLR model. A more diverse dataset is required to validate the robustness and reproducibility of the DLR model. Third, our study did not consider genetic markers. Multiple factors should be considered for more personalized treatment, including biology, pathology, genomics [24, 26, 42–52], and imaging biomarkers [53]. In addition, this study was limited to CT images despite the essentiality of MIR images in surgical planning due to their excellent resolution for soft tissues. The focus should, therefore, be given towards developing an additional model combining CT and MRI image features. Finally, the primary tumor volumes were manually delineated for feature extraction. Even though the delineations are commonly used with confirmation by another radiation oncologist in radiotherapy planning, previous studies showed that semiautomatic tumor segmentation could reduce interobserver variability and therefore is more suitable for radiomics studies [54].

5. Conclusions

This study details the development of a noninvasive predictive model that combines deep learning-based radiomics features, handcrafted features, and clinical factors to predict survival rates within three years at diagnosis of esophageal cancer. The performance of the proposed DLR nomogram is superior to the traditional radiomics model in terms of Harrel’s concordance index and AUC. The calibration curves show the good prediction performance of the nomogram. The nomogram-predicted Kaplan–Meier survival (KMS) curves differed significantly from the nonsurvival groups in the log-rank test ( value <0.05). The proposed model can present the basis for clinicians to make better treatment decisions and personalized diagnoses. Future works will include the model improvement based on larger data and complementary clinical factors.

Data Availability

The data used to support this research are included within the article.

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Acknowledgments

This study was funded by National Natural Science Foundation of China (Grant no. 72071183) and Natural Science Foundation of Shanxi Province of China for Youths (Grant no. 201801D221209).

Supplementary Materials

(I) HCR feature selection using the least absolute shrinkage and selection operator (LASSO) logistic regression model. Figure S1. HCR feature selection using the LASSO logistic regression algorithm. (a) The penalization coefficient λ in the LASSO model was tuned by the binomial deviance minimization criteria. The binomial deviance metrics (the y-axis) were plotted against log(λ) (the bottom x-axis). The top x-axis indicates the number of predictors with the given log(λ). Red dots indicate average binomial deviance for each model at the given λ. Vertical bars through the red dots show the upper and lower values of the binomial deviance. The vertical black lines represent the optimal λ, where the model provides the best fit to the data. As a result, the optimal λ of 0.02373184 was selected. (b) LASSO coefficient profiles of the 50 radiomics features. For the optimal λ, eighteen features with nonzero coefficient were selected. (II) DLR feature selection using the least absolute shrinkage and selection operator (LASSO) logistic regression model. Figure S2. DLR feature selection using the LASSO logistic regression algorithm. (a) The penalization coefficient λ in the LASSO model was tuned by the binomial deviance minimization criteria. As a result, the optimal λ of 0.0445107 was selected. (b) LASSO coefficient profiles of the 50 radiomics features. For the optimal λ, thirty three features with nonzero coefficient were selected. (Supplementary Materials)

References

P. C. Enzinger and R. J. Mayer, “Esophageal cancer,” New England Journal of Medicine, vol. 349, no. 23, pp. 2241–2252, 2003.
View at: Publisher Site | Google Scholar
H. P. Van, M. C. Hulshof, J. V. Lanschot et al., “Preoperative chemoradiotherapy for esophageal or junctional cancer,” Journal of Medical Imaging & Radiation Oncology, vol. 43, no. 2, pp. 215–219, 2012.
View at: Google Scholar
B. W. Stewart and P. Kleihues, “World cancer report,” IARC Press, Lyon, France, 2003, Report.
View at: Google Scholar
B. Kumbasar, “Carcinoma of esophagus: radiologic diagnosis and staging,” European Journal of Radiology, vol. 42, no. 3, pp. 170–180, 2002.
View at: Publisher Site | Google Scholar
R. Siegel, D. Naishadham, and A. Jemal, “Cancer statistics for hispanics/latinos, 2012,” CA: A Cancer Journal for Clinicians, vol. 62, no. 5, pp. 283–298, 2012.
View at: Publisher Site | Google Scholar
K. Tanaka, H. Miyata, K. Sugimura et al., “Negative influence of programmed death‐1‐ligands on the survival of esophageal cancer patients treated with chemotherapy,” Cancer Science, vol. 107, no. 6, pp. 726–733, 2016.
View at: Publisher Site | Google Scholar
T. N. Walsh, N. Noonan, D. Hollywood, A. Kelly, N. Keeling, and T. P. J. Hennessy, “A comparison of multimodal therapy and surgery for esophageal adenocarcinoma,” New England Journal of Medicine, vol. 335, no. 7, pp. 462–467, 1996.
View at: Publisher Site | Google Scholar
B. H. Burmeister, B. M. Smithers, V. Gebski et al., “Surgery alone versus chemoradiotherapy followed by surgery for resectable cancer of the oesophagus: a randomised controlled phase III trial,” The Lancet Oncology, vol. 6, no. 9, pp. 659–668, 2005.
View at: Publisher Site | Google Scholar
F.-L. Huang and S.-J. Yu, “Esophageal cancer: risk factors, genetic association, and treatment,” Asian Journal of Surgery, vol. 41, no. 3, pp. 210–215, 2018.
View at: Publisher Site | Google Scholar
P. Lambin, J. Zindler, B. G. L. Vanneste et al., “Decision support systems for personalized and participative radiation oncology,” Advanced Drug Delivery Reviews, vol. 109, pp. 131–153, 2017.
View at: Publisher Site | Google Scholar
H. J. W. L. Aerts, “The potential of radiomic-based phenotyping in precision medicine,” JAMA Oncology, vol. 2, no. 12, pp. 1636–1642, 2016.
View at: Publisher Site | Google Scholar
J. Song, J. Shi, D. Dong et al., “A new approach to predict progression-free survival in stage IV EGFR-mutant NSCLC patients with EGFR-TKI therapy,” Clinical Cancer Research, vol. 24, no. 15, pp. 3583–3592, 2018.
View at: Publisher Site | Google Scholar
D. Dong, L. Tang, Z.-Y. Li et al., “Development and validation of an individualized nomogram to identify occult peritoneal metastasis in patients with advanced gastric cancer,” Annals of Oncology, vol. 30, no. 3, pp. 431–438, 2019.
View at: Publisher Site | Google Scholar
B. Zhang, J. Tian, D. Dong et al., “Radiomics features of multiparametric MRI as novel prognostic factors in advanced nasopharyngeal carcinoma,” Clinical Cancer Research, vol. 23, no. 15, pp. 4259–4269, 2017.
View at: Publisher Site | Google Scholar
D. S. Kermany, M. Goldbaum, W. Cai et al., “Identifying medical diagnoses and treatable diseases by image-based deep learning,” Cell, vol. 172, no. 5, pp. 1122–1131, 2018.
View at: Publisher Site | Google Scholar
S. Wang, J. Shi, Z. Ye et al., “Predicting EGFR mutation status in lung adenocarcinoma on computed tomography image using deep learning,” European Respiratory Journal, vol. 53, no. 3, Article ID 1800986, 2019.
View at: Publisher Site | Google Scholar
S. Wang, M. Zhou, Z. Liu et al., “Central focused convolutional neural networks: developing a data-driven model for lung nodule segmentation,” Medical Image Analysis, vol. 40, pp. 172–183, 2017.
View at: Publisher Site | Google Scholar
J. Gao, Q. Jiang, Q. Jiang, B. Zhou, and D. Chen, “Convolutional neural networks for computer-aided detection or diagnosis in medical image analysis: an overview,” Mathematical Biosciences and Engineering, vol. 16, no. 6, pp. 6536–6561, Jul 15 2019.
View at: Publisher Site | Google Scholar
R. F. Thompson, V. Gilmer, C. D. Fuller, C. M. Carpenter, O. Morin, and S. AnejaW. D. Lindsay, H. J. W. L. Aerts, B. Agrimson et al., ““Artificial intelligence in radiation oncology: a specialty-wide disruptive transformation?” Radiotherapy & Oncology, vol. 129, 2018.
View at: Google Scholar
X. Zheng, Z. Yao, Y. Huang et al., “Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer,” Nature Communications, vol. 11, no. 1, p. 1236, 2020.
View at: Publisher Site | Google Scholar
K. Chaudhary, O. B. Poirion, L. Lu, and L. X. Garmire, “Deep learning-based multi-omics integration robustly predicts survival in liver cancer,” Clinical Cancer Research, vol. 24, no. 6, pp. 1248–1259, 2018.
View at: Publisher Site | Google Scholar
M. Jiang, C.-L. Li, X.-M. Luo et al., “Ultrasound-based deep learning radiomics in the assessment of pathological complete response to neoadjuvant chemotherapy in locally advanced breast cancer,” European Journal of Cancer, vol. 147, pp. 95–105, 2021.
View at: Publisher Site | Google Scholar
Y. Hu, C. Xie, H. Yang et al., “Computed tomography-based deep-learning prediction of neoadjuvant chemoradiotherapy treatment response in esophageal squamous cell carcinoma,” Radiotherapy & Oncology: Journal of the European Society for Therapeutic Radiology and Oncology, vol. 154, no. 1122–31, pp. 6–13, 2020.
View at: Publisher Site | Google Scholar
L. Zhang, D. Dong, W. Zhang et al., “A deep learning risk prediction model for overall survival in patients with gastric cancer: a multicenter study,” Radiotherapy & Oncology, vol. 150, 2020.
View at: Publisher Site | Google Scholar
R. Poplin, A. V. Varadarajan, K. Blumer et al., “Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning,” Nature Biomedical Engineering, vol. 2, no. 3, pp. 158–164, 2018.
View at: Publisher Site | Google Scholar
Y. Zhu, C. Man, L. Gong et al., “A deep learning radiomics model for preoperative grading in meningioma,” European Journal of Radiology, vol. 116, pp. 128–134, 2019.
View at: Publisher Site | Google Scholar
Z. Liu, Z. Jiang, L. Meng et al., “Handcrafted and deep learning-based radiomic models can distinguish GBM from brain metastasis,” Journal of Oncology, vol. 2021, no. 2, 10 pages, 2021.
View at: Publisher Site | Google Scholar
S. Bert-Ram, O. Kasia, S. Musib, C. J. R. Gary, and G. Vicky, “Radiomics in esophageal and gastric cancer,” Abdominal Radiology, vol. 44, no. 6, pp. 2048–2058, 2019.
View at: Publisher Site | Google Scholar
J. Encinas de la Iglesia, M. A. Corral de la Calle, G. C. Fernández Pérez, R. Ruano Pérez, and A. Álvarez Delgado, “Esophageal cancer: anatomic particularities, staging, and imaging techniques,” Radiología, vol. 58, no. 5, pp. 352–365, 2016.
View at: Publisher Site | Google Scholar
A. Fedorov, R. Beichel, J. Kalpathy-Cramer et al., “3D slicer as an image computing platform for the quantitative imaging network,” Magnetic Resonance Imaging, vol. 30, no. 9, pp. 1323–1341, 2012.
View at: Publisher Site | Google Scholar
R. T. H. M. Larue, J. E. van Timmeren, E. E. C. de Jong et al., “Influence of gray level discretization on radiomic feature stability for different CT scanners, tube currents and slice thicknesses: a comprehensive phantom study,” Acta Oncologica, vol. 56, no. 11, pp. 1544–1553, 2017.
View at: Publisher Site | Google Scholar
A. Zwanenburg, M. Vallières, M. A. Abdalah et al., “The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping,” Radiology, vol. 295, no. 2, pp. 328–338, 2020.
View at: Publisher Site | Google Scholar
A. Zwanenburg, S. Leger, L. Agolli et al., “Assessing robustness of radiomic features by image perturbation,” Scientific Reports, vol. 9, no. 1, p. 614, 2019.
View at: Publisher Site | Google Scholar
S. Fiset, M. L. Welch, J. Weiss et al., “Repeatability and reproducibility of MRI-based radiomic features in cervical cancer,” Radiotherapy & Oncology, vol. 135, pp. 107–114, 2019.
View at: Publisher Site | Google Scholar
J. J. M. van Griethuysen, A. Fedorov, C. Parmar et al., “Computational radiomics System to decode the radiographic phenotype,” Cancer Research, vol. 77, no. 21, pp. e104–e107, 2017.
View at: Publisher Site | Google Scholar
C. Sun and W. G. Wee, “Neighboring gray level dependence matrix for texture classification,” Computer Vision, Graphics, and Image Processing, vol. 23, no. 3, pp. 341–352, 1983.
View at: Publisher Site | Google Scholar
R. M. Haralick, K. Shanmugam, and I. H. Dinstein, “Textural features for image classification,” IEEE Transactions on Systems, Man, and Cybernetics, vol. SMC-3, no. 6, pp. 610–621, 1973.
View at: Publisher Site | Google Scholar
M. M. Galloway, “Texture analysis using gray level run lengths,” Computer Graphics and Image Processing, vol. 4, no. 2, pp. 172–179, 1975.
View at: Publisher Site | Google Scholar
G. Thibault, B. Fertil, and C. Navarro, “Texture indexes and gray level size zone matrix application to cell nuclei classification,” in Proceedings of the 10th International Conference on Pattern Recognition and Information Processing, pp. 140–145, Minsk, Belarus, November 2009.
View at: Google Scholar
M. Amadasun and R. King, “Textural features corresponding to textural properties,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 19, no. 5, pp. 1264–1274, 1989.
View at: Publisher Site | Google Scholar
P. Lambin, R. T. H. Leijenaar, T. M. Deist et al., “Radiomics: the bridge between medical imaging and personalized medicine,” Nature Reviews Clinical Oncology, vol. 14, no. 12, pp. 749–762, 2017.
View at: Publisher Site | Google Scholar
F. Chollet, “Xception: deep learning with depthwise separable convolutions,” 2017, https://arxiv.org/abs/1610.02357.
View at: Google Scholar
H. Hanchuan Peng, F. Fuhui Long, and C. Ding, “Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226–1238, 2005.
View at: Publisher Site | Google Scholar
R. Tibshirani, “Regression shrinkage and selection via the lasso: a retrospective,” Journal of the Royal Statistical Society: Series B, vol. 73, no. 3, pp. 267–288, 2011.
View at: Publisher Site | Google Scholar
H. Bozdogan, “Model selection and Akaike’s information criterion (AIC): the general theory and its analytical extensions,” Psychometrika, vol. 52, no. 3, pp. 345–370, 1987.
View at: Publisher Site | Google Scholar
M. H. Kutner, C. Nachtsheim, and J. Neter, Applied Linear Regression Models, McGraw-Hill/Irwin, New York, NY, USA, 2004.
A. A. Kramer and J. E. Zimmerman, “Assessing the calibration of mortality benchmarks in critical care: the Hosmer-Lemeshow test revisited,” Critical Care Medicine, vol. 35, no. 9, pp. 2052–2056, 2007.
View at: Publisher Site | Google Scholar
A. J. Vickers and E. B. Elkin, “Decision curve analysis: a novel method for evaluating prediction models,” Medical Decision Making, vol. 26, no. 6, pp. 565–574, 2006.
View at: Publisher Site | Google Scholar
P. Afshar, A. Mohammadi, K. N. Plataniotis, A. Oikonomou, and H. Benali, “From handcrafted to deep-learning-based cancer radiomics: challenges and opportunities,” IEEE Signal Processing Magazine, vol. 36, no. 4, pp. 132–160, 2019.
View at: Publisher Site | Google Scholar
The Cancer Genome Atlas Research Network, “Integrated genomic characterization of oesophageal carcinoma,” Nature, vol. 541, pp. 169–175, 2017.
View at: Google Scholar
H. Fei, W. Haojie, L. Yong et al., “SRPX2 knockdown inhibits cell proliferation and metastasis and promotes chemosensitivity in esophageal squamous cell carcinoma,” Biomedicine & Pharmacotherapy, vol. 109, 2018.
View at: Google Scholar
C. Ma, Z. Yao, Z. Yao, Q. Zhang, and X. Zou, “Quantitative integration of radiomic and genomic data improves survival prediction of low-grade glioma patients,” Mathematical Biosciences and Engineering, vol. 18, no. 1, pp. 727–744, 2021.
View at: Publisher Site | Google Scholar
J. P. O’Connor, E. O. Aboagye, J. E. Adams et al., “Imaging biomarker roadmap for cancer studies,” Nature Reviews. Clinical Oncology, vol. 14, pp. 169–186, 2017.
View at: Publisher Site | Google Scholar
C. Parmar, E. Rios Velazquez, R. Leijenaar et al., “Robust radiomics feature quantification using semiautomatic volumetric segmentation,” PLoS One, vol. 9, no. 7, Article ID e102107, 2014.
View at: Publisher Site | Google Scholar

Copyright

Copyright © 2022 Junxiu Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

PDF Download Citation

Download other formats

Order printed copies