Multimodality and temporal analysis of cervical cancer treatment response

Introduction

Cervical cancer is the fourth most common female malignancy worldwide and has been a significant global health burden. Cervical cancer is considered preventable, and early-stage detection is associated with significantly improved survival rates. Nevertheless, the disease remains a major cause of female mortality in low- and middle-income countries (1). In developed countries such as the USA, approximately 13,820 new cases of invasive cervical cancer are diagnosed annually, and 4,360 women still die from the disease (2). Concurrent chemoradiotherapy (CRT) is a mainstay treatment for cervical cancer. The treatment, however, is not personalized to individual patients despite a large observed variation in outcomes. Recent studies have discussed the importance of tailoring treatment to individual patients, incorporating novel pharmacological strategies and future perspectives to optimize therapeutic efficacy (3). Besides, recent studies have also emphasized the need for personalized treatment strategies to ensure equitable access to timely and high-quality care (4). Clinical stages, tumor histology, and positive lymph nodes have been reported as strong prognostic factors (5,6). However, clinical factors alone are insufficient to explain differences in treatment response. Non-invasive medical imaging, particularly functional imaging modalities such as apparent diffusion coefficient (ADC) magnetic resonance (MR) and positron emission tomography (PET), was used to study local tumor features. ADC quantifies the extracellular fluid compartment as a surrogate of cellularity. In highly proliferating tumors, high cellularity leads to lower ADC values due to restricted diffusion. Conversely, following CRT, necrosis may occur and result in increased ADC values as the tissue breaks down and water diffusion becomes less restricted. Nakamura et al. categorized patients into different groups and reported that the mean ADC value predicts disease recurrence by analyzing the receiver operating characteristic (ROC) curve (7). Harry et al. found a significant correlation between early treatment ADC values and the response, as well as between percentage change in ADC and treatment response (8). Similar results were reported by other investigators (9,10). [18F] fluorodeoxyglucose PET/computed tomography (¹⁸F-FDG PET/CT) measures glycolysis, which is elevated in actively metabolizing tumor cells, and is a sensitive measure for the detection of nodal or distant metastases in cervical cancer (11). A high tumor baseline FDG uptake has been correlated with worse outcomes (12). Besides basic imaging features, such as standardized uptake value and tumor volume, quantitative imaging analysis, extracting intricate first, second, and third-order imaging textures, was developed for medical images. Lucia et al. showed that radiomics features based on diffusion-weighted (DWI) magnetic resonance imaging (MRI) and PET are independent predictors of recurrence and loco-regional control with significantly higher prognostic power than clinical parameters alone (13). The same researchers further validated that high prediction accuracy of disease-free survival and locoregional control can be achieved with the combination of MR and PET radiomic features (14). Zhao et al. demonstrated that radiomic analysis of MR first-order and texture features could distinguish the early stages of cervical cancer and further assist clinical diagnosis (15).

Beyond single pre-treatment imaging analyses, researchers have investigated longitudinal changes during concurrent CRT to better understand treatment response. Salvo et al. used MRI to quantify tumor size changes (16). Huang et al. examined three sequential dynamic contrast-enhanced (DCE) MRI scans and showed that the number of low DCE voxels significantly correlated with cervical cancer treatment outcomes (17). The same authors subsequently established a universal threshold to quantify at-risk tumor voxels (18). Bowen et al. proposed an image histogram-based intensity analysis to quantify the tumor heterogeneity, which showed significant statistical changes in different modality images during treatment, supporting its potential for individualized therapy decision support (19).

Despite the promise of imaging-based biomarkers, there are several challenges with the clinical adoption of medical images for precision cervical cancer management. The correlation between images and outcomes is not always consistent. For example, pretreatment ADC was not shown to be predictive of cervical cancer response to CRT (20). In another study, FDG-PET offered no added prediction value to MR prediction (21). The complexity increases with available multiparametric MR, including perfusion, T1-, and T2-weighted information. Furthermore, PET and CT radiomics analyses can involve different feature extraction and classification methods. Moreover, longitudinal multimodal imaging studies with PET and MR may be cost-prohibitive as a routine clinical practice. To reduce the complexity and improve standardization of cervical cancer treatment response prediction, we aim to answer three questions: (I) Which imaging modality offers the best prediction of treatment response? (II) What is the optimal time point for acquiring predictive imaging studies? (III) Which radiomic features and classification methods are best suited for cervical cancer outcome prediction? We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1775/rc).

Methods

Overview of the dataset

The dataset used in this paper was obtained from the Cervical Cancer Tumor Heterogeneity (CCTH) collection in The Cancer Imaging Archive (22), including 23 cervical cancer patients collected at the University of Washington, with the most recent update released in 2023 (one patient’s data was excluded for incomplete measurement). The dataset was acquired in clinical patients with advanced stage IB–IVA cervical cancer, who were treated with standard combined radiation therapy with concurrent cisplatin-based chemotherapy (23). Functional MRIs, including T1- and T2-weighted, DCE, DWI, and post-contrast MRI, as well as FDG PET/CT, were obtained in parallel and prospectively aligned with the radiation therapy course. Imaging was performed at three time-points/radiation dose levels: before treatment start (dose 0, denoted as “Time-1”), early during the treatment course (2–2.5 weeks after treatment start/dose 20–25 Gy, denoted as “Time-2”), and at mid-treatment (4–5 weeks after treatment start/dose 45–50 Gy, denoted as “Time-3”). For brevity, combinations such as “Time-12” were used to indicate data from both Time-1 and Time-2.

Manual contours (regions of interest) of the tumor volumes, as defined by T2-weighted MRI and co-registration with PET/CT, were included in the CCTH for each case and each imaging time point. Responders or non-responders to radiotherapy were defined by 1-month post-treatment tumor volume regression of <10% residual volume from baseline. A full table of patient characteristics was provided in Table S1.

Each patient’s dataset consisted of two-dimensional (2D) slices from different modalities and a corresponding three-dimensional (3D) tumor mask. Since the image sizes vary across patients, we extracted the medical images from the tumor region with the given mask as the region of interest and then harmonized the matrix dimension to 256X256X32 for feature extraction.

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The whole workflow was summarized in Figure S1.

Feature extraction and selection

Categories of imaging features

The extracted features were divided based on orders.

Baseline features (zero-order features)

Baseline features include tumor size, max intensity and mean intensity.

First-order features

First-order features are also known as histogram-based features, including mean intensity, max intensity, variance, percentiles, skewness, and kurtosis. Detailed expressions of individual terms are shown in Appendix 1.

Second-order features (texture features)

Texture features involve the analysis of relationships between pairs of pixels or voxels. Texture features are usually computed with the Gray Level Co-occurrence Matrix (GLCM). GLCM elements are defined as in Eq. [1].

CΔx,Δy(i,j)=∑x=1n∑y=1m∏[I(x,y)=i,I(x+Δx,y+Δy)=j][1]

where “I” is the gray-level image, “i” and “j” are pixel values. “n” and “m” are the sizes of the image, (x, y) is the starting position, and (∆x, ∆y) represents the offset from starting position. In this paper, we used standard Haralick texture features, including contrast, energy, entropy, and homogeneity (24,25). The definition of each statistical term was summarized in Appendix 2.

For the 3D tumor under different modalities, we considered extracting GLCM features from individual 2D slices (GLCM2D) as well as directly from the 3D geometry (GLCM3D). GLCM2D features were extracted in four different in-plane directions, while GLCM3D features were extracted in thirteen different directions. Regarding Haralick texture features, all directions were considered symmetrically. Table 1 summarized the directions for GLCM2D and GLCM3D.

Higher-order features (Gabor filter)

Unlike zero-, first-, and second-order features that primarily capture pixel-level and local neighborhood information, higher-order features extract more abstract features. Typical higher-order features include Gabor filters, Fourier transformations, and wavelet transformations. Among different higher-order features, the Gabor filter is particularly effective in capturing texture features with frequency and direction representation (26). Gabor filter is often employed for edge detection and texture analysis, like mammogram tumor classification and architectural distortion detection. In this study, we computed the statistical features after applying Gabor filter to the original medical image, including mean intensity, max intensity, variance, skewness, and kurtosis. Specifically, we considered various Gabor filter parameters, including different sizes (10×15, 15×10, 15×15), frequencies (0.6, 0.8), and angles (0°, 45°, 90°, and 135°).

Feature selection methods

Four feature selection methods widely used in medical imaging research were employed in this study (27), including recursive feature elimination (RFE), random forest (RF), gradient boosting (GB), and principal component analysis (PCA). The prediction model was built based on logistic regression.

Results

The prediction performance was primarily measured by the area under the curve (AUC) with five-fold cross-validation. Additionally, three-fold cross-validation was employed to assess the sensitivity of prediction to variations in the data. Various types of features and their combinations were analyzed. For clarity, feature combinations were named simply; for instance, the combination of 2D GLCM features with zero-order features was labeled as “GLCM2D zero” in the table. We first compared the distributions of key variables between the development and validation datasets, and no significant differences were observed. A comparison of top 10 significant features was shown in Table S2.

Overall prediction accuracy

We predicted the treatment responder/non-responder using image features extracted from different methods, as discussed in Section “Feature extraction and selection”. Table 2 summarized the prediction accuracies of different image features extracted at varying time points. Key observations from the results include:

• GLCM-based features extracted from 2D slices had the best overall performance regarding prediction accuracy. Combining GLCM2D with zero-order or first-order features could further enhance the prediction accuracy, but the top contributing features were still from the GLCM2D based on the RFE analysis.
• For the same feature extraction method, adding more time points increased the overall prediction accuracy, while different features showed different sensitivity to new data and different trends to overfitting.
• When a single time point was used, Time-3 performed better than Time-1 and Time-2. This can be reasonably explained by the fact that tumor response to treatment is more evident at later time points, reflecting the patient’s response.
• When two time points were used, combinations that included Time-3 (Time-13 and Time-23) generally outperformed those without it (Time-12), except for the Gabor filter. The relative performance of Time-13 vs. Time-23 depended on the feature selection method. This pattern could be similarly explained by tumor response to treatment.

It is important to note that the reported accuracy values were obtained using all features extracted from the medical images. A detailed model complexity analysis was presented in Appendix 3. In addition, an example of prediction performance comparison between training and testing data using Time-3 and GLCM feature was provided in Table S4, with the corresponding training/testing AUC comparison shown in Table S5.

Prediction with fewer modalities

We further extended our study to evaluate the prediction accuracy using a single modality, including PET/CT (denoted as PET), DCE, and DWI-ADC (denoted as ADC), as well as the combinations of any two modalities. Prediction accuracy using all three time points was summarized in Table 3 with five-fold validation. The results showed that GLCM2D-based features had better prediction accuracy than other feature types, and combining GLCM2D with zero-order features slightly outperformed combinations with first-order features. This conclusion was consistent with the results obtained using all modality images, as reported in Section “Overall prediction accuracy”.

Moreover, we saw an overall advantage of using ADC for single modality prediction and DCE&PET for dual modality prediction. An exception was observed with DCE-based Gabor zero features, which initially yielded the highest prediction accuracy in five-fold validation, though this result was inconsistent with other data in the same column. To verify this anomaly, we conducted an additional three-fold validation, as shown in Table 4. The new results indicated that DCE-based Gabor zero features no longer produced the highest accuracy, while other features that performed well in the five-fold validation remained reliable. Therefore, we concluded that GLCM2D-based features showed greater stability, and were generally superior to other feature types.

Understanding the prediction power of GLCM features

We further applied feature selection to reduce model complexity and assess the significance of individual features using RFE, PCA, GB, and RF. Figure 1 shows the AUCs for varying numbers of features extracted from ADC across all three time points. We considered feature numbers from 3 to 20 due to limited data size. The results showed relatively stable performance even when the number of features was reduced to three.

Figure 1 Prediction accuracy from different feature extraction methods. AUC, area under the curve; GB, gradient boosting; PCA, principal component analysis; RF, random forest; RFE, recursive feature elimination.

To understand the prediction power of individual GLCM features, we chose the ADC GLCM feature, randomly selected the top 5, 12, and 20 features, and summarized the names of features in Table 5. Among different types of GLCM features, contrast was the most significant feature, and correlation also played an important part in certain cases, while energy and homogeneity were not significant parameters in determining treatment effectiveness.

Discussion

Here, we studied image-based outcome prediction for cervical cancer patients undergoing CRT. Our paper has explored several vital questions, including the prediction effectiveness of different modalities and modalities measured at different treatment stages.

Prediction results with different imaging modalities

Previous studies explored either the prediction with a single modality (7,8,11) or multiple modalities (13,19). There has not been a consensus as to which imaging modality is best for cervical cancer outcome prediction (28); our analysis has shown that for the three individual modalities considered, ADC has the highest prediction accuracy, followed by PET/CT, while DCE has the lowest prediction accuracy. When considering dual modalities, the combination of DCE and PET/CT has the highest prediction accuracy.

In addition, our GLCM-based method shows that contrast and correlation are the two most important factors in predicting the treatment response, which is consistent with the fact that MRI provides superior soft tissue contrast (29).

Our study found that ADC better predicts than DCE for cervical cancer treatment response, which indicates the significance of cellularity in tumor aggressiveness and response to therapy. This observation is consistent with previous reports (30,31). Though DCE has been reported to enhance detectability of residual cancer (32,33), there are no existing findings validating the superior of DCE Gabor filter features over the texture features. When utilizing dual modalities, we found the combination of DCE and PET to be slightly superior for response prediction due to their complementary information (34,35).

Prediction results with fewer time points

When evaluating prediction using a single time point, imaging data measured at Time-3 have better performance over other individual time points, including the combination of time points 1 and 2. This is expected as Time-3 reflects the post-treatment tumor status. A similar conclusion is reported in laryngeal carcinoma evaluation (36). Unfortunately, imaging information at this time point cannot be used to adapt the treatment plan that has already been delivered. When adding more time points, the prediction accuracy generally increases compared to a single time point. Adding more time points helps establish the temporal dynamics of the tumor responding to treatment (37). However, in this study, compared to using all three time points, the prediction accuracy of Time-3 only decreases by around 4% with GLCM2D features and decreases by around 9% with GLCM2D with zero-order features, indicating minimal contribution to the prediction with the temporal information.

Feature extraction methods

Compared to the previous study (19) which uses the first-order features, our study evaluates a broader range of feature extraction methods and demonstrates that the GLCM features hold a clear advantage for the cervical cancer data considered. We attribute this to the GLCM features’ robustness. Since GLCM considers the relative spatial relationship between pixels, it tends to be less sensitive to variation in image acquisition parameters among patients and image modalities. Although adding zero and first-order features could be complementary to texture features, texture features always play a dominant role in determining treatment effectiveness. This observation is consistent with other radiotherapy research, such as in lung tumors, bladder tumors, and glioblastoma (38-41). Moreover, for individual GLCM features extracted from the images, our study shows that Contrast is the most predictive GLCM feature, while energy and homogeneity are not significant in determining the treatment response. The results indicate that the contrast of the tumor is more indicative of its response to CRT than its energy and homogeneity. In comparison, the uniformity and concentration of pixels reflected by energy and homogeneity are relatively insensitive to reflecting the underlying tumor aggressiveness and response, which is consistent with pulmonary tumor analysis (42-44).

Prediction methods

Besides the comparison of feature extraction methods, another contribution of this study is the comparison of feature selection methods. In this study, we show that the RFE and RF are more effective than other feature extraction methods. RFE has been proven to be an efficient way to select features in different tumor research, like tumor classifications and predictions in cervical cancer (19), prostate cancer (45), and lung cancer (46). Similarly, RF has been successfully applied to different types of medical image classification, labeling, and segmentation (47).

RFE-based approaches have demonstrated superiority, particularly in medical applications where the number of features far exceeds the number of samples (48,49). Unlike PCA, which transforms original features into a new set of uncorrelated components, or decision-tree-based methods like RF and GB, RFE focuses on selecting the most critical features for prediction, significantly simplifying the feature extraction process. However, RFE’s focus on feature selection can make it more susceptible to overfitting, especially with smaller datasets or when the feature selection process is overly aggressive. In contrast, decision-tree-based methods like RF and GB are more complex but offer greater robustness against overfitting. In this study, RF slightly outperformed RFE, though their relative performance may vary depending on the specific dataset.

Limitations and future work

One limitation of this study is the absence of detailed clinical and treatment background information, which may serve as important variables. Although our preliminary analysis indicated no significant improvement when incorporating the limited clinical features, future research should explore multi-source models that integrate radiomics with comprehensive clinical and treatment data to enhance predictive accuracy and clinical relevance.

Another limitation is the relatively small cohort size, which may impact the robustness and generalizability of the findings. To address this, future studies should involve larger, multi-institutional longitudinal datasets to validate the performance and stability of radiomic features in predicting cervical cancer outcomes.

Beyond direct extensions of this work, the framework of longitudinal radiomic analysis can be applied to other tumor types to support personalized treatment planning and response monitoring. Expanding this approach across different cancers may further demonstrate the potential of radiomics in precision oncology.

Conclusions

This paper explores how multimodality images (PET/CT, ADC, DCE) at different treatment points can affect the prediction of treatment response for cervical cancer. Our research explores different image feature extraction methods and feature selection methods to understand the best prediction approach.

The key conclusions and contributions of this paper are:

• We discover that GLCM-based features are more effective compared to other types of image features considered regarding prediction accuracy. Moreover, 2D GLCM features are more effective when predicting the treatment response compared to 3D GLCM. We further discover that within the 2D GLCM features, Contrast and Correlation are the main contributors to the prediction from RFE extraction.
• We compare the prediction effectiveness between different treatment stages. Our results have shown that post-stage (Time-3) measurement has the highest prediction accuracy than pre-stage (Time-1) and mid-stage (Time-2). When using GLCM features, post-stage measurement is only 4% lower than using all time points with 2D GLCM features and 9% lower than using 2D GLCM with zero-order features, indicating minimal contribution to the prediction with the temporal information.
• We further compare the prediction accuracy between different modalities and their combinations. We discover that ADC has the best prediction performance compared to PET/CT, followed by DCE with GLCM-based features. Also, combining DCE and PET/CT is more effective compared to the other two possible combinations. These results not only provide a comprehensive comparison of the relationship between different modalities and time points, but also indicate the potential to reduce the measurements and enhance cervical cancer treatment efficiency.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1775/rc

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1775/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. LaVigne AW, Triedman SA, Randall TC, Trimble EL, Viswanathan AN. Cervical cancer in low and middle income countries: Addressing barriers to radiotherapy delivery. Gynecol Oncol Rep 2017;22:16-20. [Crossref] [PubMed]
2. American Cancer Society. Cancer Facts & Figures 2024. Atlanta: American Cancer Society; 2024. Available online: https://www.cancer.org/research/cancer-facts-statistics/all-cancer-facts-figures/2024-cancer-facts-figures.html
3. D'Oria O, Bogani G, Cuccu I, D'Auge TG, Di Donato V, Caserta D, Giannini A. Pharmacotherapy for the treatment of recurrent cervical cancer: an update of the literature. Expert Opin Pharmacother 2024;25:55-65. [Crossref] [PubMed]
4. D'Augè TG, Di Donato V, Giannini A. Strategic approaches in management of early-stage cervical cancer: A comprehensive editorial. Clin Exp Obstet Gynecol 2024;51:235.
5. Liu YM, Ni LQ, Wang SS, Lv QL, Chen WJ, Ying SP. Outcome and prognostic factors in cervical cancer patients treated with surgery and concurrent chemoradiotherapy: a retrospective study. World J Surg Oncol 2018;16:18. [Crossref] [PubMed]
6. Liu J, Tang G, Zhou Q, Kuang W. Outcomes and prognostic factors in patients with locally advanced cervical cancer treated with concurrent chemoradiotherapy. Radiat Oncol 2022;17:142. [Crossref] [PubMed]
7. Nakamura K, Joja I, Nagasaka T, Fukushima C, Kusumoto T, Seki N, Hongo A, Kodama J, Hiramatsu Y. The mean apparent diffusion coefficient value (ADCmean) on primary cervical cancer is a predictive marker for disease recurrence. Gynecol Oncol 2012;127:478-83. [Crossref] [PubMed]
8. Harry VN, Persad S, Bassaw B, Parkin D. Diffusion-weighted MRI to detect early response to chemoradiation in cervical cancer: A systematic review and meta-analysis. Gynecol Oncol Rep 2021;38:100883. [Crossref] [PubMed]
9. Qin F, Pang H, Ma J, Xu H, Yu T, Luo Y, Dong Y. The value of multiparametric MRI combined with clinical prognostic parameters in predicting the 5-year survival of stage IIIC1 cervical squamous cell carcinoma. Eur J Radiol 2023;169:111181. [Crossref] [PubMed]
10. Holopainen E, Lahtinen O, Könönen M, Anttila M, Vanninen R, Lindgren A. Greater increases in intratumoral apparent diffusion coefficients after chemoradiotherapy predict better overall survival of patients with cervical cancer. PLoS One 2023;18:e0285786. [Crossref] [PubMed]
11. Calles-Sastre L, Mucientes-Rasilla J, San-Frutos Llorente LM, Royuela A, Garcia-Espantaleón Navas M, Herrero Gámiz S, Pérez-Medina T. Prognostic significance of metabolic tumor volume and total lesion glycolysis in patients with advanced cervical carcinoma. Rev Esp Med Nucl Imagen Mol (Engl Ed) 2019;38:17-21. [Crossref] [PubMed]
12. Onal C, Guler OC, Torun N, Reyhan M. Long-term assessment of clinical parameters and positron emission tomography parameters in predicting recurrence in uterine cervical cancer patients receiving definitive chemoradiotherapy. Nucl Med Commun 2024;45:203-10. [Crossref] [PubMed]
13. Lucia F, Visvikis D, Desseroit MC, Miranda O, Malhaire JP, Robin P, Pradier O, Hatt M, Schick U. Prediction of outcome using pretreatment (18)F-FDG PET/CT and MRI radiomics in locally advanced cervical cancer treated with chemoradiotherapy. Eur J Nucl Med Mol Imaging 2018;45:768-86. [Crossref] [PubMed]
14. Lucia F, Visvikis D, Vallières M, Desseroit MC, Miranda O, Robin P, Bonaffini PA, Alfieri J, Masson I, Mervoyer A, Reinhold C, Pradier O, Hatt M, Schick U. External validation of a combined PET and MRI radiomics model for prediction of recurrence in cervical cancer patients treated with chemoradiotherapy. Eur J Nucl Med Mol Imaging 2019;46:864-77. [Crossref] [PubMed]
15. Zhao X, Wang X, Zhang B, Liu X, Xuan D, Xia Y, Zhang X. Classifying early stages of cervical cancer with MRI-based radiomics. Magn Reson Imaging 2022;89:70-6. [Crossref] [PubMed]
16. Salvo G, Odetto D, Saez Perrotta MC, Noll F, Perrotta M, Pareja R, Wernicke A, Ramirez PT. Measurement of tumor size in early cervical cancer: an ever-evolving paradigm. Int J Gynecol Cancer 2020;30:1215-23. [Crossref] [PubMed]
17. Huang Z, Mayr NA, Lo SS, Grecula JC, Wang JZ, Jia G, Yuh WT. Characterizing at-Risk Voxels by Using Perfusion Magnetic Resonance Imaging for Cervical Cancer during Radiotherapy. J Cancer Sci Ther 2012;4:254-9. [Crossref] [PubMed]
18. Huang Z, Yuh KA, Lo SS, Grecula JC, Sammet S, Sammet CL, Jia G, Knopp MV, Wu Q, Beauchamp NJ 3rd, Yuh WT, Wang R, Mayr NA. Validation of optimal DCE-MRI perfusion threshold to classify at-risk tumor imaging voxels in heterogeneous cervical cancer for outcome prediction. Magn Reson Imaging 2014;32:1198-205. [Crossref] [PubMed]
19. Bowen SR, Yuh WTC, Hippe DS, Wu W, Partridge SC, Elias S, Jia G, Huang Z, Sandison GA, Nelson D, Knopp MV, Lo SS, Kinahan PE, Mayr NA. Tumor radiomic heterogeneity: Multiparametric functional imaging to characterize variability and predict response following cervical cancer radiation therapy. J Magn Reson Imaging 2018;47:1388-96. [Crossref] [PubMed]
20. Meyer HJ, Wienke A, Surov A. Pre-treatment Apparent Diffusion Coefficient Does Not Predict Therapy Response to Radiochemotherapy in Cervical Cancer: A Systematic Review and Meta-analysis. Anticancer Res 2021;41:1163-70. [Crossref] [PubMed]
21. van Kol K, Ebisch R, Beugeling M, Cnossen J, Nederend J, van Hamont D, Coppus S, Piek J, Bekkers R. Comparing Methods to Determine Complete Response to Chemoradiation in Patients with Locally Advanced Cervical Cancer. Cancers (Basel) 2023;16:198. [Crossref] [PubMed]
22. Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M, Tarbox L, Prior F. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging 2013;26:1045-57. [Crossref] [PubMed]
23. Mayr N, Yuh WT, Bowen S, Harkenrider M, Knopp MV, Lee E, Leung E, Lo S, Small Jr. W, Wolfson H. Cervical Cancer - Tumor Heterogeneity: Serial Functional and Molecular Imaging Across the Radiation Therapy Course in Advanced Cervical Cancer (Version 1). The Cancer Imaging Archive 2023. Available online: https://www.cancerimagingarchive.net/collection/cc-tumor-heterogeneity/
24. Haralick RM. Statistical and Structural Approaches to Texture. Proceedings of the IEEE 1979;67:786-804.
25. Feng H, Subramaniyan SP, Tewani H, Prabhakar P. Physics-Constrained Neural Network for design and feature-based optimization of weave architectures. Composites Part A: Applied Science and Manufacturing 2024;187:108465.
26. Gabor D. Theory of communication. Part 1: The analysis of information. Electrical Engineers Part III Radio & Communication Engineering Journal of the Institution of 1946;93:429–41.
27. Chowdhary CL, Acharjya DP. Segmentation and Feature Extraction in Medical Imaging: A Systematic Review. Procedia Computer Science 2020;167:26-36.
28. Mayerhoefer ME, Prosch H, Beer L, Tamandl D, Beyer T, Hoeller C, et al. PET/MRI versus PET/CT in oncology: a prospective single-center study of 330 examinations focusing on implications for patient management and cost considerations. Eur J Nucl Med Mol Imaging 2020;47:51-60. [Crossref] [PubMed]
29. Cleary JO, Guimarães AR. Magnetic Resonance Imaging. In Pathobiology of Human Disease 2014;3987-4004.
30. Kim HS, Kim CK, Park BK, Huh SJ, Kim B. Evaluation of therapeutic response to concurrent chemoradiotherapy in patients with cervical cancer using diffusion-weighted MR imaging. J Magn Reson Imaging 2013;37:187-93. [Crossref] [PubMed]
31. Fu ZZ, Peng Y, Cao LY, Chen YS, Li K, Fu BH. Value of apparent diffusion coefficient (ADC) in assessing radiotherapy and chemotherapy success in cervical cancer. Magn Reson Imaging 2015;33:516-24. [Crossref] [PubMed]
32. Huang JW, Song JC, Chen T, Yang M, Ma ZL. Making the invisible visible: improving detectability of MRI-invisible residual cervical cancer after conisation by DCE-MRI. Clin Radiol 2019;74:166.e15-21. [Crossref] [PubMed]
33. Yuh WT, Mayr NA, Jarjoura D, Wu D, Grecula JC, Lo SS, Edwards SM, Magnotta VA, Sammet S, Zhang H, Montebello JF, Fowler J, Knopp MV, Wang JZ. Predicting control of primary tumor and survival by DCE MRI during early therapy in cervical cancer. Invest Radiol 2009;44:343-50. [Crossref] [PubMed]
34. Besson FL, Fernandez B, Faure S, Mercier O, Seferian A, Mignard X, et al. 18F-FDG PET and DCE kinetic modeling and their correlations in primary NSCLC: first voxel-wise correlative analysis of human simultaneous [18F]FDG PET-MRI data. EJNMMI Res 2020;10:88.
35. De Bruyne S, Van Damme N, Smeets P, Ferdinande L, Ceelen W, Mertens J, Van de Wiele C, Troisi R, Libbrecht L, Laurent S, Geboes K, Peeters M. Value of DCE-MRI and FDG-PET/CT in the prediction of response to preoperative chemotherapy with bevacizumab for colorectal liver metastases. Br J Cancer 2012;106:1926-33. [Crossref] [PubMed]
36. Ljumanovic R, Langendijk JA, Hoekstra OS, Knol DL, Leemans CR, Castelijns JA. Pre- and post-radiotherapy MRI results as a predictive model for response in laryngeal carcinoma. Eur Radiol 2008;18:2231-40. [Crossref] [PubMed]
37. Nardone V, Reginelli A, Grassi R, Boldrini L, Vacca G. D”Ippolito E, Annunziata S, Farchione A, Belfiore MP, Desideri I, Cappabianca S. Delta radiomics: a systematic review. Radiol Med 2021;126:1571-83. [Crossref] [PubMed]
38. Scalco E, Rizzo G. Texture analysis of medical images for radiotherapy applications. Br J Radiol 2017;90:20160642. [Crossref] [PubMed]
39. Xu X, Zhang X, Tian Q, Zhang G, Liu Y, Cui G, Meng J, Wu Y, Liu T, Yang Z, Lu H. Three-dimensional texture features from intensity and high-order derivative maps for the discrimination between bladder tumors and wall tissues via MRI. Int J Comput Assist Radiol Surg 2017;12:645-56. [Crossref] [PubMed]
40. Liu Y, Xu X, Yin L, Zhang X, Li L, Lu H. Relationship between Glioblastoma Heterogeneity and Survival Time: An MR Imaging Texture Analysis. AJNR Am J Neuroradiol 2017;38:1695-701. [Crossref] [PubMed]
41. Brynolfsson P, Nilsson D, Torheim T, Asklund T, Karlsson CT, Trygg J, Nyholm T, Garpebring A. Haralick texture features from apparent diffusion coefficient (ADC) MRI images depend on imaging and pre-processing parameters. Sci Rep 2017;7:4041. [Crossref] [PubMed]
42. Kim H, Park CM, Lee M, Park SJ, Song YS, Lee JH, Hwang EJ, Goo JM. Impact of Reconstruction Algorithms on CT Radiomic Features of Pulmonary Tumors: Analysis of Intra- and Inter-Reader Variability and Inter-Reconstruction Algorithm Variability. PLoS One 2016;11:e0164924. [Crossref] [PubMed]
43. Pyka T, Bundschuh RA, Andratschke N, Mayer B, Specht HM, Papp L, Zsótér N, Essler M. Textural features in pre-treatment [F18]-FDG-PET/CT are correlated with risk of local recurrence and disease-specific survival in early stage NSCLC patients receiving primary stereotactic radiation therapy. Radiat Oncol 2015;10:100.
44. Cook GJ, Yip C, Siddique M, Goh V, Chicklore S, Roy A, Marsden P, Ahmad S, Landau D. Are pretreatment 18F-FDG PET tumor textural features in non-small cell lung cancer associated with response and survival after chemoradiotherapy? J Nucl Med 2013;54:19-26. [Crossref] [PubMed]
45. Fehr D, Veeraraghavan H, Wibmer A, Gondo T, Matsumoto K, Vargas HA, Sala E, Hricak H, Deasy JO. Automatic classification of prostate cancer Gleason scores from multiparametric magnetic resonance images. Proc Natl Acad Sci U S A 2015;112:E6265-73. [Crossref] [PubMed]
46. Hage Chehade A, Abdallah N, Marion JM, Oueidat M, Chauvet P. Lung and colon cancer classification using medical imaging: a feature engineering approach. Phys Eng Sci Med 2022;45:729-46. [Crossref] [PubMed]
47. Criminisi A, Shotton J, editors. Decision Forests for Computer Vision and Medical Image Analysis. London: Springer Science & Business Media; 2013.
48. Bolón-Canedo V, Sánchez-Maroño N, Alonso-Betanzos A. A review of feature selection methods on synthetic data. Knowledge and information systems 2013;34:483-519.
49. Remeseiro B, Bolon-Canedo V. A review of feature selection methods in medical applications. Comput Biol Med 2019;112:103375. [Crossref] [PubMed]