Variations in radiomic features of the femoral head and neck during helical tomotherapy in prostate and rectal cancer patients

Abstract

Background

This study introduces a novel approach for screening bone alterations in the femoral head and neck (H&N) of prostate cancer (PCa) and rectal cancer (RCa) patients during helical tomotherapy (HT) using megavoltage computed tomography (MVCT) images. The goal is to identify robust radiomic features (RFs) with the highest percentage changes during treatment and examine their correlation with the administered dose.

Methods

Reproducible RFs were identified through a test-retest analysis using a cheese phantom. This study involved 20 male patients (10 PCa, 10 RCa). The left and right femoral H&N regions were segmented on MVCT images from the initial, middle, and final HT sessions. Absolute RF values and relative percentage changes were analyzed using repeated measures analysis of variance (ANOVA). The Pearson correlation coefficient with Benjamini-Hochberg adjustment (q < 0.05) was used to evaluate the relationship between altered RFs and the dose (Gy).

Results

The femoral H&N had the highest relative percentage changes (RPCs) for intensity-histogram (IH) and intensity-based (IB) RFs, respectively, with texture-based RFs providing insights into radiotherapy-induced changes. Strong correlations (r ~ -0.7) were observed between changes in H&N RFs and dose (Gy) in PCa. For RCa, IB features, including the IB_Coefficient_of_Variation (r = 0.54) for the neck, and IH RFs, such as IH_Minimum_Histogram_Gradient (r = -0.51) and IB_Robust_Mean_Absolute_Deviation (r = 0.55) for the head, showed significant correlations.

Conclusions

Among robust RFs, the most highly correlated with dose alterations were IB, IH, and gray-level co-occurrence matrix (GLCM)-based RFs. The RFs at mid- and end-treatment varied with dose fractionation. Percentage changes in robust MVCT features during HT may serve as early markers.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-14903-5.

Background

Prostate cancer (PCa) ranks as the second most common cancer among men worldwide, with external beam radiation therapy (EBRT) serving as an effective treatment modality for many patients [1, 2]. Colorectal cancer (CRC) is equally prevalent, ranking as the third most common cancer and the third leading cause of cancer-related deaths in both men and women. Rectal cancer (RCa) accounts for about 12% of these cases [3]. For RCa patients, EBRT—whether administered as a postoperative adjunct or preoperative neoadjuvant therapy—plays a crucial role in achieving treatment success. Given that over half of cancer patients undergo EBRT, either as a standalone treatment or combined with surgery and chemotherapy, optimizing this therapy is essential [4].

Helical tomotherapy (HT), a sophisticated form of intensity-modulated radiotherapy (IMRT), represents an advanced approach by integrating continuous helical motion with megavoltage computed tomography (MVCT) for real-time patient positioning verification [5, 6]. A significant advantage of HT over conventional IMRT is its ability to deliver an optimized and homogeneous dose to the tumor while minimizing radiation exposure to the organs at risk (OARs) [6].

In recent years, modern Monte Carlo (MC)–based dose calculation algorithms have been increasingly recognized for their superior accuracy in heterogeneous regions, such as bone–soft tissue interfaces, which are highly relevant to the femoral head and neck (H&N). Although many developments have focused on linear accelerators, the implementation of MC algorithms in treatment planning systems also provides valuable opportunities for tomotherapy systems, potentially improving dose calculation accuracy in regions prone to skeletal toxicity [7].

Despite these advancements, radiation-induced late toxicity—particularly bone toxicity—remains a significant concern for patients receiving pelvic radiation therapy (RT), which can result from both direct and indirect effects of radiation on bone tissues [8, 9]. Clinical studies have shown that cancer patients undergoing EBRT are at an increased risk of developing osteoporosis and experiencing pathological fractures due to accelerated loss of bone mineral density (BMD) associated with treatment. Notably, radiation-related toxicity in the femoral H&N region of the femur, such as fractures, cortical bone thinning, and necrosis, may manifest months or even years after radiotherapy [10, 11].

Treating fractures in irradiated bones presents significant challenges due to the increased risk of morbidity and mortality. Fractures resulting from RT often have a high rate of non-union [12, 13]. Even with successful treatment, patients frequently experience chronic pelvic and back pain, as well as reduced mobility, which significantly impacts their quality of life [14, 15]. While the mechanisms of radiation-induced bone damage are not fully understood, studies have highlighted key effects. Radiotherapy can alter the biomechanical properties of bones, such as causing matrix embrittlement and suppressing normal osteoblast proliferation [16, 17]. Additionally, radiation has been shown to increase the sensitivity of bone cells to apoptosis [18]. Efforts are ongoing to establish pre-therapeutic parameters for early therapy response and prognosis assessment, including late bone toxicity, which could guide treatment modifications and improve patient survival [10, 19]. The focus is on identifying biomarkers that allow early intervention and enhance therapeutic outcomes, though such biomarkers are not yet available.

Recent developments in image analysis and biomarker research have led to the creation of advanced tools for diagnosing diseases, predicting outcomes, and managing treatments [20]. Radiomics is a state-of-the-art image-processing paradigm developed to identify biomarkers by extracting high-throughput features from medical images. Considerable research has been conducted on the use of radiomic features (RFs) extracted from CT images for the detection, diagnosis, evaluation, and automatic classification of bone diseases, including osteoporosis [19, 21]. Additionally, the radiomic literature has demonstrated correlations between alterations in imaging features, radiation dose, and clinical effects [22–25]. A feature extraction algorithm converts image data from a region of interest (ROI) into first-order and higher-order RFs. By analyzing the inherent relationships among various levels of clinical data, radiomics has the potential to enhance diagnostic accuracy and offer prognostic and predictive insights [26].

Texture analysis (TA) is a mathematical approach that evaluates grayscale intensities and pixel locations within an image, allowing for the quantification of subtle image inhomogeneities that may be imperceptible to the naked eye. This method generates textural features (TF), which are derived from analysis-based methods [19]. The texture of an image refers to the spatial arrangement of pixels with varying intensities. Texture measures can quantify gray-level variations, capturing repetitive patterns and uniformity in the image pixels [27]. These texture parameters have been utilized for analyzing trabecular bone microstructure in computed tomography (CT) scans [28]. The LIFEx package, which complies with the Image Biomarker Standardization Initiative (IBSI), is one of the radiomics software tools used for this analysis [29–31].

There is a notable research gap, as no studies have examined the changes in RFs in the femoral H&N regions during HT using treatment planning MVCT images. Beyond its role in patient positioning, MVCT provides a unique opportunity for real-time monitoring of bone changes during HT. Unlike conventional imaging, MVCT enables repeated acquisition throughout the treatment course without requiring additional imaging sessions, allowing for the dynamic assessment of bone structural alterations in response to radiation. To our knowledge, no prior studies have leveraged MVCT to track changes in the femoral H&N during HT, highlighting the novelty of this approach. Integrating MVCT-based monitoring with radiomic analysis may facilitate early detection of bone toxicity and provide a non-invasive biomarker for treatment adaptation. Specifically, this study aims to assess the changes in reproducible RFs of the femoral H&N in patients with PCa and RCa undergoing HT and to identify correlations between these changes and the administered radiation dose.

Methods

Study design and workflow implementation

The study workflow, outlined in Fig. 1, involved two key steps: (I) a test-retest analysis using a cheese phantom, and (II) an analysis of reproducible RFs with dose in PCa and RCa patients. Each step included a radiomic workflow, comprising data gathering, image segmentation, image preprocessing, feature extraction, and statistical analysis. The subsequent sections provide detailed explanations of these steps.

Fig. 1.

Study workflow: (I) Test-retest analysis using a cheese phantom. (II) Analysis of reproducible RFs with dose in PCa and RCa patients. Steps include data gathering, segmentation, preprocessing, feature extraction, and statistical analysis (concordance correlation coefficient (CCC) in I, Pearson correlation in II)

I. Test-retest analysis using a cheese phantom for selection of robustness features

To ensure robustness of feature identification, a cheese phantom was used for the test-retest analysis (Fig. 1, I). This process involved imaging the cheese phantom, ROIs delineation, image preprocessing, feature extraction, and statistical analysis. The following sections elaborate on each step in detail.

Cheese Phantom image acquisition

A cylindrical TomoTherapy “cheese” phantom, also referred to as a tomo-phantom or Sun Nuclear Tomo-Phantom, was employed to extract RFs from various materials (Fig. 2). This phantom is commonly used for quality assurance in HT and is designed to mimic the radiation attenuation properties of water. The manufacturer specifies the electron density of the Sun Nuclear Tomo-Phantom HE as 1.000 ± 0.005 relative to water. The phantom is a cylindrical blue object with a diameter of 300 mm and a length of 180 mm, featuring several holes. Half of the Tomo-Phantom features four ball bearings (BB) fiducials embedded on its surface within the central transverse slice. In the present study, bone plugs with different densities were utilized: inner bone (1.148 g/cm³), CB2-30% (1.333 g/cm³), CB2-50% (1.56 g/cm³), and cortical bone (1.821 g/cm³).

Fig. 2.

(A) Bone plugs within cheese phantom, (B) The cheese phantom on the helical tomotherapy couch

Two sets of test and retest MVCT images were acquired for the cheese phantom, with a 15-minute interval between scans involving setup changes. The MVCT protocol was consistent for both scans, employing the standard algorithm, normal pitch, and 4 mm interval.

Phantom image segmentation

ROIs were manually delineated in the initial image and replicated in the second image to ensure consistency in the ROI for feature extraction, maintaining identical volume. These contours were segmented and visualized using different colors in the LIFEx package v7.4.0, as illustrated in Fig. 3 [29].

Fig. 3.

ROIs highlighted in various colors corresponding to the bone plugs on the cheese phantom MVCT image in the LIFEx software

Phantom image preprocessing

After image segmentation and prior to feature extraction, image preprocessing was performed to homogenize the images, reduce noise, increase sensitivity, and normalize the intensities across the two images. This preprocessing was consistent for both images. The intensity rescaling method used was relative (ROI: min < > max), with image intensities discretized to 400 Gy levels and 10 size bins. Spatial resampling was conducted considering spacing X, Y, and Z to be 1 × 1 × 1 mm³.

Phantom image feature extraction

After delineation of the ROI, a total of 94 RFs were extracted for each region. These RFs were classified into first-order and second-order categories. First-order features were further divided into intensity-based (IB) and intensity-histogram (IH) RFs. IB features were derived from the intensity values, while IH features are statistical features derived from histograms of the intensity values.

Second-order features, also known as textural features, are calculated based on the spatial relationships between voxels. These features are further classified into four categories: gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), and neighboring gray-tone difference matrix (NGTDM) features.

Phantom statistical analysis

Following feature extraction, the concordance correlation coefficient (CCC) was computed based on the two sets of images. Features with a CCC of less than 0.9 were considered non-repeatable and excluded from further analysis [32–34].

II. Reproducible feature-dose analysis in patients with prostate and rectal cancers

After acquiring reproducible RFs based on the cheese phantom test-retest analysis in the previous section, the next main step, according to Fig. 1, panel B, which involves the following five steps: patient data collection (MVCT images and HT radiation doses), patient image segmentation, patient image preprocessing, patient image feature extraction, and patient statistical analysis. The following sections elaborate on each step in detail.

Patient data collection

From January 2019 to December 2022, a retrospective selection of 20 patients at our institution was conducted, including 10 with PCa and 10 with RCa. The study was approved by the Ethics Committee of the Iran University of Medical Sciences under ethics approval number IR.IUMS.FMD.REC.1401.387. Additionally, the Institutional Review Board approved the authorization for this retrospective analysis. All relevant clinical data and dosimetric information from the patient’s medical records were collected for the present study. MVCT images from the initial, middle, and final sessions were collected. Patients were included based on the visibility of the entire femoral H&N in their MVCT scans. Moreover, patients with metallic prostheses were excluded because artifacts from these implants can interfere with the accurate delineation of the femoral H&N and affect radiomic feature extraction.

None of the patients with PCa received chemotherapy as part of their treatment regimen. In contrast, patients with RCa underwent chemotherapy with Capecitabine (150 and 500 mg/m², twice daily for 5 days per week) throughout their radiotherapy course. Patients underwent HT with individualized treatment plans based on their specific treatment objectives. PCa patients received doses of 6900, 7200, and 7820 cGy across 32, 30, and 34 fractions, respectively, while RCa patients received 4500, 5000, 5040, and 5600 cGy across 25, 25, 28, and 28 fractions, respectively.

Radiotherapy was administered using a tomotherapy linear accelerator (Radixact X9) with 6 MeV photon beams. The radiation dose is delivered helically in HT technology through 51 projections per rotation. Dynamic multi-leaf collimator IMRT is delivered on a continuous helix and utilizes an integrated MVCT unit that allows for verification of patient positioning. DICOM MVCT files and dose information were exported from the treatment planning archive (Precision workstation). The prescribed radiation dose varied according to their treatment regimen.

Radiation dose data for the patients were obtained from the dose-volume histogram (DVH) values in the radiotherapy plan report. While the doses for the left and right femur differ, they are identical for the H&N on each side.

Patient image segmentation

An experienced radiologist manually delineated four regions, including the right and left femoral sections of both H&N, on the transaxial slices of initial, middle, and final MVCT images of HT sessions using the LIFEx package V7.4.0 [29]. The goal was to maintain a fixed volume for each region in each patient, and three-dimensional (3D) acquired VOIs were employed for further feature extraction and analysis. As illustrated in Fig. 4, the femoral areas are displayed on transaxial slices (panels A and B) and a coronal slice (panel C) using different colors: blue for the right femoral head, pink for the left femoral head, green for the right femoral neck, and yellow for the left femoral neck.

Fig. 4.

Segmentation of a PCa patient: (A) Transaxial slice with left and right femoral heads segmentation. (B) Transaxial view of four four-segmented femoral H&N regions. (C) Coronal view of the same four regions

Patient image preprocessing

Following the manual segmentation, image pre-processing was conducted using the LIFEx package v7.4.0 [29]. Pre-processing encompassed spatial resampling, intensity discretization, and intensity rescaling, with consistent settings applied across all patients. Preprocessing parameters were the same as the ones for the cheese phantom to be independent of its effects.

Patient feature extraction

After preprocessing the MVCT images of PCa and RCa patients undergoing HT, the same set of RFs was extracted as those acquired from the MVCT images of the cheese phantom. The robust features derived from the test-retest exam were analyzed. Consequently, only the stable features from the initial, middle, and final session MVCTs were retained for further analysis.

The RFs were extracted in six different conditions: (1–3) absolute RFs from the MVCT images of initial, middle, and final treatments, and (4–6) three sets of relative RFs based on changes from MVCT images of initial to middle treatments, initial to final treatments, and middle to final treatments. These features are labeled with _0, _1, _2, _01, _02, and _12 at the end of the feature name to ensure distinct identifiers, respectively.

Patient statistical analysis

Initially, the normality of the data was assessed using the Shapiro-Wilk test to determine whether to use Pearson correlation or Spearman rank correlation analysis. To identify meaningful differences among features from three time points: the initial, middle, and final sessions, a repeated-measure analysis of variance (ANOVA) was conducted using SPSS software v26.0. A p-value of less than 0.05 is considered statistically significant, indicating that there is likely a difference between the various treatment time points.

Relative percentage change (RPC) was calculated as the difference between the mean of the second measurement and the mean of the first measurement, relative to the mean of the first measurement (_02 RFs) as the most prominent set of relative RFs compared to _01 and _12 RFs. The RPC analysis was conducted for four regions—left femur head, left femur neck, right femur head, and right femur neck —across patients with PCa and RCa during HT. Moreover, the Pearson correlation coefficient, followed by Benjamini and Hochberg p-value correction, was calculated between both absolute and relative RFs, and the administered dose from the final session, considering q-value < 0.05 was significant. This dose is approximately twice that of the middle session, making any subsequent analysis consistent whether the dose from the final or middle session is used.

Results

Phantom statistical analysis

Out of the 94 extracted features, 73 were identified as robust based on their CCC (greater than 0.95) during the analysis of bone plugs (see Supplementary Table 1 for the list of all features). These robust features were subsequently used in the statistical analysis to assess radiomic changes in MVCT images of both PCa and RCa patients.

Table 1.

Features (Bold letters) with the relative changes in pca. RPC values are shown in parentheses

Features	Femoral Head		Femoral Neck
	Left	Right	Left	Right
IB	None	−10th Percentile (−15%)	- Coefficient of Variation (23%) −10th Percentile (51%) −25th Percentile (177%) −50th Percentile (24%) - Mean (−13%) - Median (−24%) - Quartile Coefficient of Dispersion (33%)	- Coefficient of Variation (18%) −10th Percentile (−29%) −25th Percentile (−12%) −50th Percentile (−22%) - Mean (−13%) - Median (−22%) - Quartile Coefficient of Dispersion (%29)
IH	Variance (21%)	- Maximum Histogram Gradient (13%) - Minimum Histogram Gradient (12%) - Minimum Histogram Gradient Grey Level (23%)	None	None
GLCM	Joint Maximum (−17%)	None	None	None
GLRLM	- Long Run Low Grey Level Emphasis (24%) - Low Grey Level Run Emphasis (25%) - Short Run Low Grey Level Emphasis (25%)	None	- Long Run Low Grey Level Emphasis (8%) - Low Grey Level Run Emphasis (8%) - Short Run Low Grey Level Emphasis (8%)	- Long Run Low Grey Level Emphasis (23%) - Low Grey Level Run Emphasis (23%) - Short Run Low Grey Level Emphasis (23%)
NGTDM	- Complexity (15%) - Strength (13%)	None	None	None

Patient statistical analysis: relative percentage changes

Tables 1 and 2 summarize the RFs with the most relative percentage changes (RPCs) indicated in parentheses for patients with PCa and RCa, respectively. Among the 73 reproducible RFs analyzed, only those with RPC values higher than 5% are listed, while RFs with an RPC of less than 5% are not included in the tables, despite all 73 RFs showing alterations.

Table 2.

Features with the relative changes in rca. RPC values are shown in parentheses

Features	Femoral Head		Femoral Neck
	Left	Right	Left	Right
IB	None	None	- Coefficient of Variation (20%) −10th Percentile (−56%) −25th Percentile (−21%) −50th Percentile (−20%) −75th Percentile (−11%) - Mean (−15%) - Median (−20%) - Energy (−12%) - Quartile Coefficient of Dispersion (23%) - Mode (18%) - Minimum Histogram Gradient Grey Level (−15%)	- Coefficient of Variation (21%) −10th Percentile (−48%) −25th Percentile (−119%) −50th Percentile (−20%) −75thPercentile (−11%) - Mean (−15%) - Median (−20%) - Energy (−12%) - Quartile Coefficient of Dispersion (25%) - Mode (−11%) - Minimum Histogram Gradient Grey Level (−15%)
IH	- Variance (10%) - Minimum Histogram Gradient Grey Level (12%) - Coefficient of Variation (11%) - Quartile Coefficient Of Dispersion (14%)	- Minimum Histogram Gradient Grey Level (10%)	None	None
GLCM	None	- Angular Second Moment (11%) - Joint Maximum (34%)	Joint Maximum (18%)	Joint Maximum (12%)
GLRLM	- Long Run Low Grey Level Emphasis (15%) - Low Grey Level Run Emphasis (15%) - Short Run Low Grey Level Emphasis (15%)	None	- Long Run Low Grey Level Emphasis (18%) - Low Grey Level Run Emphasis (18%) - Short Run Low Grey Level Emphasis (18%)	- Long Run Low Grey Level Emphasis (30%) - Low Grey Level Run Emphasis (30%) - Short Run Low Grey Level Emphasis (30%)
NGTDM	- Complexity (15%) - Strength (10%)	Complexity (−11%)	None	None

In PCa patients (Table 1), the femoral neck displayed the highest RPCs in IB features, with the 10th percentile showing the largest changes of 51% in the left neck and 29% in the right. The Quartile Coefficient of Dispersion also demonstrated substantial increases on both sides of the neck, with 33% on the left and 29% on the right. In contrast, the femoral head exhibited fewer IB changes, the most significant being a 15% decrease in the 10th percentile on the left side. IH features showed alterations only in the femoral head, with no changes observed in the neck. In the left femoral head, the only significant change was observed in Variance (21%). For the right femoral head, the Minimum Histogram Gradient Grey Level (23%) exhibited the greatest alteration. Joint Maximum (−17%) is the only GLCM feature observed in the left femoral head. GLRLM features exhibited alterations in Long-Run and Short-Run metrics across all regions, except for the right femoral head, with changes generally ranging from 23 to 25%. NGTDM features, including Complexity (15%) and Strength (10%), were altered in the left femoral head.

In PCa, IB and GLRLM features showed alterations in the neck regions, whereas IH, GLCM, GLRLM, and NGTDM features changed in the femoral head. This highlights the significance of these features in their respective regions.

In RCa patients (Table 2), the femoral neck exhibited the highest RPCs in IB features, with the 10th percentile showing a 56% change in the left neck and the 25th percentile showing a 119% change in the right neck. In the IB category, the 75th Percentile, Energy, Mode, and Minimum Histogram Gradient Grey Level features, which were not observed in PCa, were seen in the femoral neck regions of RCa patients. IH features showed alterations only in the femoral head, with no changes observed in the neck. In the GLCM category, Joint Maximum exhibited the most significant changes, with a 34% increase in the right femoral head, an 18% increase in the left femoral neck, and a 12% increase in the right femoral neck. In RCa patients, GLRLM features exhibited alterations across all regions except for the right femoral head, with changes generally ranging from 15 to 30%, consistent with the pattern observed in PCa patients. NGTDM features, including Complexity (15%) and Strength (10%), were altered in the left femoral head, with no significant changes observed in the neck. This pattern in the left femoral head is similar to that seen in PCa; however, in RCa, Complexity also showed changes in the right femoral head.

In RCa, alterations were observed in IB, GLCM, and GLRLM features in the neck regions. Conversely, IH, GLCM, GLRLM, and NGTDM features exhibited changes in the femoral head. This underscores the relevance of these features in their specific anatomical locations.

Patient statistical analysis: repeated measure ANOVA

Tables 3 and 4 present RFs in PCa and RCa patients, respectively, highlighting significant differences for the femoral H&N at three time points. According to Table 3, for IB features, several metrics, including the 10th Percentile, 25th Percentile, and Mean, exhibited notable changes, highlighting the dynamic alterations in bone characteristics over time. In contrast, no significant changes were detected in the femoral head. IH features exhibited significant differences in the femoral head, particularly in Median Absolute Deviation, Standard Deviation, Variance, and Minimum Histogram Gradient Grey Level. In contrast, the femoral neck showed alterations exclusively in the right region. No significant changes were observed in the GLCM and GLRLM categories across any regions. Additionally, NGTDM features demonstrated significant changes in the femoral left neck, with Coarseness and Strength showing notable differences. These findings underscore the importance of focusing on the femoral neck in PCa for detecting RF changes and suggest that the femoral head may not exhibit the same degree of temporal variation in these features. In PCa, IB, IH, and NGTDM features exhibited alterations in the neck regions, while changes in the femoral head were observed only for IH features.

Table 3.

The significant differences at three time points for the femoral H&N in PCa patients

Features	Femoral head		Femoral neck
	Left	Right	Left	Right
IB	None	None	−10th Percentile (p = 0.01) −25th Percentile (p = 0.02) −50th Percentile (p = 0.01) −90th Percentile (p = 0.048) - Coefficient of Variation (p = 0.03) - Mean (p = 0.01) - Mean Absolute Deviation (p = 0.02) - Median (p = 0.006) - Median Absolute Deviation (p = 0.02) - Standard Deviation (p = 0.01) - Variance (p = 0.015)	−10th Percentile (p = 0.02) −25thPercentile (p = 0.02) −50thPercentile (p = 0.003) −75thPercentile (p = 0.01) −90thPercentile (p = 0.047) - Mean (p = 0.02) - Median (p = 0.003) - LIB1 Global Intensity Peak(0.5mL) (p = 0.03) - LIB Global Intensity Peak(1mL) (p = 0.02)
IH	- Median Absolute Deviation (p = 0.049) - Standard Deviation (p = 0.04) - Variance (p = 0.048)	Minimum Histogram Gradient Grey Level (p = 0.008)	None	- LIH2 Global Intensity Peak(0.5mL) (p = 0.03) - LIH Global Intensity Peak(1mL) (p = 0.02)
GLCM	None	None	None	None
GLRLM	None	None	None	None
NGTDM	None	None	- Coarseness (p = 0.001) - Strength (p = 0.04)	None

¹Local Intensity-Based

²Local Intensity-Histogram

Table 4.

The significant differences at three time points for the femoral head and neck in RCa patients

Features	Femoral head		Femoral neck
	Left	Right	Left	Right
IB	−10thPercentile (p = 0.01) −25thPercentile (p = 0.02) −50thPercentile (p = 0.02) −90thPercentile (p = 0.046) - Coefficient of Variation (p = 0.02) - Mean (p = 0.03) - Median (p = 0.02) - Quartile Coefficient of Dispersion (p = 0.02)	−25thPercentile (p = 0.01) −50thPercentile (p = 0.04) −90thPercentile (p = 0.049) - Coefficient of Variation (p = 0.03) - Mean (p = 0.03) - Median (p = 0.04) - Energy (p = 0.048) - Quartile Coefficient of Dispersion (p = 0.02)	−10thPercentile (p = 0.003) −25thPercentile (p < 0001) −50thPercentile (p = 0.002) −75thPercentile (p < 0001) −90thPercentile (p = 0.008) - Coefficient of Variation (p < 0001) - Mean (p < 0001) - Median (p = 0.002) - Energy (p = 0.01) - Quartile Coefficient of Dispersion (p < 0001) - Robust Mean Absolute Deviation (p = 0.03) - Root Mean Square (p = 0.005)	−10thPercentile (p < 0001) −25thPercentile (p < 0001) −50thPercentile (p = 0.002) −75thPercentile (p = 0.003) −90thPercentile (p = 0.006) - Coefficient of Variation (p < 0001) - Mean (p = 0.002) - Median (p = 0.002) - Energy (p = 0.007) - Quartile Coefficient of Dispersion (p < 0001) - Root Mean Square (p = 0.005) - Total Calcium Score (p = 0.01)
IH	None	None	None	−25th Percentile (p = 0.01) −50th Percentile (p = 0.02) −75th Percentile (p = 0.02) −90th Percentile (p = 0.006) - Coefficient of Variation (p = 0.01) - Mean (p = 0.02) - Median (p = 0.016) - Quartile Coefficient of Dispersion (p = 0.02)
GLCM	None	- Joint Average (p = 0.02) - Sum Average (p = 0.02)	None	None
GLRLM	None	None	None	None
NGTDM	None	None	None	None

In RCa, significant alterations in IB features were observed across the femoral H&N regions. IH RFs displayed substantial variations only in the femoral right neck. No significant alterations were detected in the IH RFs of the femoral head. GLCM features showed significant differences in the right femoral head, specifically in Joint Average and Sum Average. No significant changes were noted in the GLRLM and NGTDM categories across any regions.

Patient statistical analysis: pearson correlation

Figure 5 presents a composite heatmap for PCa, combining six heatmaps for feature categories _0, _1, _2, _01, _02, and _12. In this heatmap, red denotes a perfect positive correlation, blue signifies a perfect negative correlation and other colors represent varying levels of correlation. Darker shades (categories _0, _1, and _2) indicate more variable correlations with both positive and negative relationships, suggesting diverse interactions among absolute RFs. In contrast, lighter shades (categories _01, _02, and _12) show more consistent and stronger correlations, reflecting more stable RF relationships in the relative category.

Fig. 5.

Pearson self and cross-correlation between radiation dose and absolute RFs (_0, _1, and _2) and relative RFs (_01, _02, and _12) in PCa patients

Also, Fig. 6 illustrates a similar composite heatmap for RCa. The patterns are comparable to those in PCa: darker categories (_0, _1, and _2) exhibit variable correlations, while lighter categories (_01, _02, and _12) show more uniform relationships. Both cancers display these trends, with darker patterns indicating complex interactions and lighter patterns suggesting more stable correlations.

Fig. 6.

Pearson self and cross-correlation between radiation dose and absolute RFs (_0, _1, and _2) and relative RFs (_01, _02, and _12) in RCa patients, with red indicating positive correlations and blue indicating negative correlations

Supplementary Figs. 1 to 12 present heatmaps for each feature category of both cancer types, highlighting RFs with Pearson correlations greater than 0.4 or less than − 0.4 with the dose.

Figure 7 shows Pearson correlations of H&N RFs with the dose (Gy) for relative categories of _01 and _02 RFs in PCa. Panel A: IB_90th Percentile01_H (−0.78); Panel B: IB_75thPercentile01_N (−0.71); Panel C: LIB_Global Intensity Peak(1mL)_02_H (−0.76); Panel D: IB_75th Percentile02_N (−0.61). These features have the highest correlations with the dose (Gy), indicating strong relationships.

Fig. 7.

Pearson correlations of head and neck RFs with the dose (Gy) for relative categories _01 and _02 RFs in PCa patients, showing strong relationships. Panels: (A) IB_90thPercentile01_H, (B) IB_75thPercentile01_N, (C) LIB_GlobalIntensityPeak(1mL)_02_H, and (D) IB_75thPercentile02_N

Figures 8 and 9 each contain four panels for relative features in RCa. In both figures, panels A and B display the highest positive Pearson correlations of the _01 and _02 features with the dose (Gy) in the H&N region, respectively. In Fig. 8, panels C and D show the two highest negative correlations in the neck region, while no negative correlation in the femoral head exceeds 0.4 or falls below − 0.4. Figure 9C illustrates the _02 features with the highest negative correlation in the neck, and panel D presents the only relative _12 RFs in the neck region with the highest negative Pearson correlation with dose (Gy). Positive correlations in the H&N region suggest sensitivity to dose increases, aiding in dose-related tracking. Significant negative correlations highlight an inverse relationship with dose (Gy), reflecting treatment effects. The modest correlations in the femoral H&N regions indicate a less pronounced impact of dose, emphasizing the need to consider anatomical differences.

Fig. 8.

(A, B) Highest positive and (C, D) Highest negative Pearson correlations of the relative _01 RFs in H&N with the dose (Gy) in RCa patients

Fig. 9.

(A, B) Highest positive Pearson correlations of the relative _02 RFs in the femoral H&N regions, respectively. (C) The highest negative correlation of the _02 RFs in the femoral head. (D) The highest negative correlation of relative _12 RF in the femoral head, all concerning dose (Gy) in RCa patients

These results in Figs. 7, 8 and 9 underscore that these specific RFs have the highest correlations with dose (Gy), though additional RFs also demonstrate significant correlations, as depicted in Supplementary Figure Heatmaps 1 to 6 for PCa and 7 to 12 for RCa. These supplementary figures provide further insights into the broader spectrum of RFs that are responsive to dose (Gy) variations. Relative RF_12 exhibited the lowest correlations compared to other groups.

Discussion

In this study, we demonstrated for the first time the feasibility of utilizing RF extracted from MVCT images to assess radiotherapy-induced alterations in the femoral H&N among PCa and RCa patients who underwent HT. Additionally, the test-retest cheese phantom was utilized to identify RFs that are consistent and reliable across multiple scans under identical conditions. High reproducibility in these features strengthens their potential to enhance radiomic models, improving diagnostic accuracy, prognostic evaluation, and treatment response in clinical practice. Radiomics offers a quantitative and objective approach to enhancing cancer detection and treatment decisions [35, 36].

By extracting detailed information from medical images and integrating it with clinical, genomic, and other data, radiomics can identify biomarkers crucial for diagnosis and therapy [24, 35, 36]. This field leverages advanced computational techniques to address clinical challenges [26, 37, 38]. In radiomic studies, each feature represents a specific aspect of tissue characteristics, and variations in these feature values can reflect underlying structural and functional changes in the tissues being analyzed [38, 39]. For instance, according to Tables 1 and 2, GLCM Joint Maximum is the only RF in the left femoral head of PCa patients with a significant RPC, while the same RF shows meaningful RPCs in all three other regions of RCa patients. This feature represents the frequency of the most predominant pair of neighboring intensity values. A higher GLCM Joint Maximum indicates that a particular intensity pair occurs frequently in the image, suggesting greater uniformity or repetitive texture patterns.

Moreover, the GLRLM category in Tables 1 and 2, including Long Run Low Grey Level Emphasis, Low Grey Level Run Emphasis, and Short Run Low Grey Level Emphasis, showed significant RPCs in both PCa and RCa patients. These features capture texture patterns by analyzing the distribution of consecutive pixels with the same gray level. The observed RPCs in these features across the left and right femoral regions suggest notable alterations in tissue texture during treatment, potentially reflecting changes in the underlying tissue structure and composition. Furthermore, within the NGTDM category, complexity and strength are RFs with meaningful RPC values in both cancers, exhibiting a modest Pearson correlation with administered dose, as shown in Supplementary Fig. 5 for PCa and Figs. 7 and 8, and 11 for RCa. The observed RPCs in these RFs suggest notable changes in the complexity and uniformity of tissue texture during treatment.

While there is no study specifically utilizing MVCT images to evaluate RFs of the femoral H&N, the literature includes a few similar studies investigating RF changes in bone due to radiation. These studies, although employing different methodologies and designs, share the common goal of assessing radiation-induced alterations in bone. Abdollahi et al. [2] examined RF changes in the femoral head of PCa patients using magnetic resonance (MR) images before and three months after IMRT. Significant alterations were found in the IH and GLCM features. Notably, a decrease in the energy feature indicated increased tissue heterogeneity, while changes in the contrast feature suggested substantial local variations due to radiation [2]. In this study, as shown in Table 1, IH and GLCM features exhibited changes only in the femoral head, consistent with Abdollahi et al., as their study focused solely on this region. However, while Abdollahi et al. reported significant alterations in energy and contrast, no significant changes in these features were observed in this study. Additionally, unlike their study, we extended our analysis to the femoral neck, where no notable changes were observed in IH and GLCM features.

Hayar et al. [3] analyzed bone density changes in the L5 vertebra, sacrum, and femoral head before and three months after RT in RCa patients, using both KVCT and dual-energy X-ray absorptiometry (DEXA) imaging. They found significant changes in bone density in regions assessed by KVCT, whereas DEXA revealed no substantial change in the femoral heads [3]. Yaprak et al. [40] conducted a study to assess bone density and fractures in the L1 and L2 vertebrae using Hounsfield units (HU) calculated from CT of the treatment planning system. After one year, they analyzed HU percentage changes in 57 patients, noting that 4 patients experienced fractures. Their findings indicated that fractures were more common in patients with lower HU values [40].

Nardone et al. [19] assessed the utility of TA on CT simulation images for predicting pelvic fractures in patients undergoing radiotherapy. Using LIFEx software, they evaluated images of the L5 vertebra, sacrum, and femoral heads. Fracture patients were matched 1:1 with control patients who did not develop fractures. The study concluded that 3D-bone CT TA can effectively stratify the risk of radiation-induced pelvic fractures. The TA variables that resulted significantly in binary logistic regression were L5-energy and FH-skewness [19]. However, the skewness feature was removed during our test-retest phantom analysis as redundant, while the energy feature, classified as an IB_RF, showed significant alterations in both the RPC and ANOVA tests for RCa patients.

In line with our findings, Ishikawa et al. [41] reported that pretreatment CT bone density was the only independent predictor of posttreatment fractures in cervical cancer patients receiving RT. While dose–volume parameters were associated with fracture risk in univariate analysis, multivariate modeling confirmed bone density as the dominant factor. This highlights that structural bone quality, likely reflecting osteoporosis, plays a central role in determining fracture risk, consistent with our observation that IB- and IH-based features, which are sensitive to density alterations, showed the greatest changes during treatment. Together, these results suggest that radiomic features extracted from MVCT may provide a non-invasive surrogate for bone density assessment, potentially allowing for earlier identification of patients at high risk for insufficiency fractures.

Similarly, Rydzewski et al. [21] investigated rib fracture risk after stereotactic body radiation therapy (SBRT) for early-stage non-small cell lung cancer by integrating dosimetric parameters, radiomic features, and a bone density score derived from HUs. They demonstrated that dosimetric variables were the strongest predictors of fracture risk, but that bone density and radiomics improved risk stratification, with their combined model achieving an area under the curve of 0.864. These results reinforce the clinical utility of radiomics as an adjunct to traditional dosimetric and bone density measures for predicting fracture risk. In line with this, the current study shows that MVCT-derived radiomic features can capture dose-related changes in pelvic bone structure during treatment, suggesting their potential role as early imaging biomarkers to complement post-treatment fracture risk models.

According to Tables 3 and 4, the findings from the repeated measures ANOVA indicate that RT induces significant changes in bone texture and density in the femoral H&N, with potential lateralization in response due to anatomical or functional differences. Additionally, the differences in the specific imaging features between PCa and RCa patients may be influenced by the difference in their treatment courses, with chemotherapy in RCa patients possibly intensifying or altering radiation-induced bone changes compared to PCa patients. Regarding first-order RFs, specifically IB and IH, as shown in Table 3, most RFs exhibiting significant changes in the right and left femoral necks in the PCa cohort were associated with IB features, whereas in the right H&N and left femoral head, the significant changes were predominantly linked to IH features. In contrast, Table 4 reveals differing results for the RCa, where significant features in both the femoral H&N were primarily IB features, with some IH features in the right femoral neck also demonstrating meaningful differences.

Based on the repeated-measure ANOVA results detailed in Table 3, the NGTDM_Coarseness and NGTDM_Strength RFs exhibited significant p-values in PCa, indicating meaningful alterations in texture. NGTDM_Strength was also captured as a significant feature in the RPC test of PCa patients. Strength measures the uniformity of gray levels within the neighborhood of a pixel. It quantifies the degree to which similar gray tones dominate a region. Coarseness specifically quantifies the amount of local variation in gray levels within a specified neighborhood around each pixel. For RCa in Table 4, the GLCM_Joint Average and GLCM_Sum Average RFs were significant, reflecting notable changes in texture characteristics. Joint Average measures the average intensity of all pixel pairs within the GLCM, reflecting the overall gray level of the image’s texture. The Sum Average quantifies the mean of the summed gray levels of pixel pairs, giving insight into the overall intensity distribution in the image.

To further clarify the biological and structural meaning of the most informative radiomic features, we highlight the key RFs identified in our study. NGTDM_Complexity quantifies the variability in gray levels within a defined neighborhood and reflects how heterogeneous the local bone texture is; increases in this feature during treatment may indicate disrupted trabecular structure or altered bone microarchitecture. NGTDM_Strength measures the dominance of uniform gray levels across neighboring pixels, serving as an indicator of tissue uniformity; decreases in this feature suggest reduced homogeneity and possible early marrow changes. GLCM_Joint Maximum represents the highest frequency of co-occurring gray-level pairs, with higher values indicating repetitive or uniform texture patterns; variations in this feature during therapy may signal changes in trabecular density or mineralization patterns. First-order features from the IB and IH categories capture statistical characteristics of voxel intensities, such as mean, variance, and skewness, which are closely related to bone mineral content and microarchitecture. Observed changes in these features across treatment time points may reflect early bone marrow fat infiltration, decreased trabecular connectivity, or other dose-related structural alterations. Collectively, these features provide interpretable imaging biomarkers that link quantitative texture and intensity patterns with underlying radiotherapy-induced bone changes, supporting their potential utility for early detection of bone toxicity and fracture risk.

One of the study objectives was to compare HU-based first-order RFs, which are derived from statistical measures of HU such as mean, variance, skewness, and kurtosis, with second-order texture features to determine which can detect changes earlier. ANOVA analysis indicated that the observed changes in both cancer types were primarily associated with IB_RFs and IH_RFs, suggesting that first-order categories are more sensitive to variations in imaging features across the treatment period. The greater number of significant changes in IB and IH features points to alterations in bone mineral content and microarchitecture, which could have clinical implications for the long-term management of bone health in cancer patients undergoing RT.

Figures 5 and 6 show that darker categories (absolute RFs: _0, _1, and _2) exhibit variable correlation patterns, suggesting complex interactions among features. In contrast, lighter categories (relative RFs: _01, _02, and _12) display more consistent and stronger correlations, indicating more stable relationships. These stable features might be more reliable for predicting treatment outcomes and potential bone toxicity. The consistent correlations in lighter categories could be useful for identifying patients at risk before therapy begins, while the variability in darker categories may reflect more complex feature interactions. This distinction helps prioritize features that can effectively predict treatment responses and guide personalized interventions. RFs in both absolute and relative states were analyzed to assess their relationship with dose (Gy). The observed relationship suggests that these RFs could be utilized to detect bone insufficiencies earlier using MVCT during treatment, potentially allowing for preemptive intervention before complications arise.

This study uniquely leverages MVCT images to track radiotherapy-induced changes in the femoral H&N during HT. Unlike CT or MRI, which are typically acquired months after treatment and offer limited insights, MVCT images are readily available before and during HT as part of the treatment planning process, providing real-time data for ongoing assessment. Importantly, our study did not involve patients with fractures or bone complications; instead, it focused on monitoring RF changes across courses of HT to evaluate the feasibility of tracking radiation-induced alterations. Through test-retest analysis, reproducible RFs were identified, and subsequent ANOVA and RPC analyses revealed features that significantly change during therapy. Pearson correlation analysis was also employed to determine the relationship between RFs and the administered dose, further elucidating the dose-response dynamics.

The systematic review by Berk [42] highlights that radiation can induce trabecular thinning, decreased bone mineral density, and alterations in bone turnover, ultimately increasing fracture risk. These findings align with our observations of MVCT-derived radiomic features, particularly IB, IH, GLCM, and NGTDM metrics, which capture changes in bone density, heterogeneity, and microarchitecture during treatment. Rasmusson et al. [1] reported that while EBRT for PCa did not significantly increase hip fractures, it was associated with a higher incidence of severe complications such as hip arthroplasties, underscoring the need for long-term surveillance and preventive strategies. Recent literature underscores the significant impact of high-dose radiotherapy on bone structure and function. Sakyi et al. [43] further demonstrated in a rat model that EBRT leads to reduced bone volume fraction and trabecular thickness in the femoral neck and lumbar spine, resulting in decreased bone strength, emphasizing the importance of monitoring skeletal integrity. Collectively, these studies support the potential of integrating MVCT-based radiomic biomarkers with clinical insights to predict structural bone changes, enable early interventions, and mitigate long-term skeletal complications.

Strong negative correlations (approximately − 0.7) were found between changes in H&N RFs and dose (Gy) in PCa. In contrast, for RCa patients, significant correlations were observed with IB features, such as the IB Coefficient of Variation (r = 0.54) for the neck, and IH features, including the IH Minimum Histogram Gradient (r = −0.51) and IB_Robust Mean Absolute Deviation (r = 0.55) for the head. These variations in association with administered dose may reflect different pathophysiological mechanisms underlying these features with increasing dose, potentially related to femoral head artery necrosis, increased bone marrow fat, or decreased trabecular network connectivity, necessitating further histopathological studies. However, the results of this study demonstrate that low-resolution MVCT images can be useful for monitoring bone loss during treatment, complementing existing literature that assesses bone alterations post-treatment. As the discussion of feature changes and their relationship to pathophysiology is a novel area of study, it is important to define and explore these aspects further in the near future. In summary, IB and IH-based features, which exhibit a negative correlation with dose (Gy), may indicate an increase in bone marrow fat and a decrease in signal intensity. Conversely, changes in texture features such as GLCM and NGTDM might serve as biomarkers for the loss of trabecular network connectivity.

While the current study uniquely demonstrates the feasibility of MVCT radiomics for real-time monitoring of bone changes during HT, it is important to consider the broader context of multimodality imaging [44]. Many recent radiomics and machine learning (ML) studies integrate multiple imaging modalities, such as positron emission tomography (PET)/CT and single photon emission computed tomography (SPECT)/CT, to improve robustness through complementary information from metabolic and anatomical data [45, 46]. For example, in PCa and RCa, [¹⁸F]FDG PET/CT–based radiomics has been shown to enhance predictive modeling by enabling cross-modality feature selection and correlation analysis [46]. Although the current study focused on MVCT as a single modality, these findings suggest that future research may benefit from integrating MVCT with additional imaging modalities to further strengthen predictive power and clinical applicability.

Beyond monitoring radiation-induced changes in the femoral H&N, MVCT-based radiomics has demonstrated prognostic potential in other clinical contexts. For example, a recent study in H&N squamous cell carcinoma showed that radiomic features extracted from MVCT images acquired during RT could stratify patients by overall survival, with predictive performance comparable to planning CT–based radiomics and superior to clinical factors alone [47]. This finding suggests that MVCT radiomics can provide real-time, clinically relevant biomarkers beyond bone assessment, reinforcing its broader applicability in treatment monitoring and personalized therapy planning.

Despite the significance of our findings, several limitations must be acknowledged. First, the small sample size limits generalizability, even when the robust features were applied, but future studies should include larger cohorts. Second, the absence of a detailed clinical history regarding bone health prevented us from comparing RFs between patients with healthy bones and those with bone complications. Establishing distinct patient groups based on bone health status in future research could provide insights into the differences between normal and compromised bone features. Third, patients with metallic prostheses were excluded because artifacts from these implants interfere with accurate delineation of the femoral H&N and can distort the gray-level intensities in MVCT images, potentially leading to unreliable radiomic feature extraction. The presence of prostheses could artificially alter texture and first-order features, introduce noise, and confound the assessment of radiation-induced bone changes, thereby affecting the robustness and reproducibility of the results. Future studies may explore artifact-reduction techniques or alternative imaging strategies to include patients with prostheses and expand the applicability of MVCT radiomics. Fourth, the single-center data collection limits the applicability of our results; multi-center studies would provide more generalizable insights. Fifth, while the CCC method was utilized for feature selection, more advanced methods could be employed in future research to enhance clinical relevance, potentially incorporating ML techniques for improved clinical applications.

Sixth, one limitation of this study is the reliance on manual segmentation, which is time-consuming and subject to inter- and intra-observer variability. Future work could benefit from the integration of automated or semi-automated segmentation methods—particularly those based on deep learning—to enhance consistency, reduce labor intensity, and enable large-scale clinical application [48–50]. Advancing this field will require the adoption of explainable artificial intelligence methods, enabling radiomics-based models to deliver interpretable insights that can directly inform clinical decision-making [51, 52]. Finally, the absence of long-term follow-up is a notable limitation, as radiation-induced femoral head damage often develops years after treatment. For future studies, performing quantitative computed tomography before and after treatment is recommended to validate these extracted features by correlating them with bone mineral density. Nonetheless, our study highlights the potential of early RF analysis during tomotherapy as a valuable tool for predicting post-radiotherapy bone fractures and other complications, warranting further investigation in future studies.

Conclusions

Of the 73 robust features extracted from MVCT images during tomotherapy, only 12—primarily from IB, IH, and GLCM categories—demonstrated strong correlations with dose-related changes in the femoral H&N. The alterations in features observed at mid-treatment and the end of treatment varied depending on dose fractionation and the time points analyzed. Moreover, percentage relative changes in robust features during tomotherapy have the potential to serve as non-invasive imaging biomarkers and early radiomic markers from MVCT images, offering promise for predicting future bone complications following radiotherapy. Specific features such as NGTDM_Complexity, NGTDM_Strength, and GLCM_Joint Maximum may provide valuable insights into radiotherapy-induced femoral H&N bone changes and further enlighten osseous pathologic processes. Early detection of structural changes may help mitigate the risk of fractures in patients with PCa and RCa following HT for future study.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ANOVA

Analysis of variance

BB

Ball bearing

BMD

Bone mineral density

CCC

Concordance correlation coefficient

DEXA

Dual-energy X-ray absorptiometry

DVH

Dose-volume histogram

EBRT

External beam radiation therapy

GLCM

Gray-level co-occurrence matrix

GLRLM

Gray-level run length matrix

H&N

Head and neck

HT

Helical tomotherapy

IB

Intensity-Based

IH

Intensity-histogram

IMRT

Intensity-modulated radiotherapy

LIB

Local intensity-based

LIH

Local intensity-histogram

MC

Monte Carlo

ML

Machine learning

MR

Magnetic resonance

MVCT

Megavoltage computed tomography

NGTDM

Neighboring gray-tone difference matrix

OARs

Organs at risk

PET

Photon emission computed tomography

PCa

Prostate cancer

RCa

Rectal cancer

RF

Radiomic features

RMSE

Root-mean-squared error

ROIs

Regions of interest

RPC

Relative percentage change

RT

Radiation therapy

SBRT

Stereotactic body radiation therapy

SPECT

Single photon emission computed tomography

TA

Texture analysis

TF

Texture features

VOI

Volume of interest

Author contributions

Guarantors of the integrity of the entire study: S.R. Mahdavi, M. Malekzadeh. Study conception and design: S.R. Mahdavi, M. Malekzadeh, M. Gholizade, E. Yazdani. Literature review: M. Gholizade, E. Yazdani, M. Malekzadeh. Clinical investigations: M. Gholizade, A. Nikoofar, F. Goli-Ahmadabad, G. Esmaili •Experimental work and data analysis: M. Gholizade, S.R. Mahdavi, M. Malekzadeh, E. Yazdani, F. Goli-Ahmadabad, G. Esmaili •Statistical analysis: F.S. Hosseini-Baharanchi, M. Gholizade, E. Yazdani. Manuscript drafting: E. Yazdani, M. Gholizade. Critical revision of the manuscript: S.R. Mahdavi, M. Malekzadeh, F.S. Hosseini-Baharanchi. All authors read and approved the final manuscript and agree to be personally accountable for their contributions and to ensure the accuracy and integrity of the work.

Funding

This work was supported by the School of Medicine, Iran University of Medical Sciences (IUMS), and Grant number 24666.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to ethical restrictions and the need to protect patient confidentiality, following institutional and data protection regulations. However, they are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The Ethics Committee of the Iran University of Medical Sciences (IUMS) approved this retrospective study under ethics approval code IR.IUMS.FMD.REC.1401.387. All methods were carried out in accordance with relevant guidelines and regulations. The requirement for informed consent was waived by the Ethics Committee, as the study involved retrospective analysis of anonymized data obtained from routine clinical practice.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.