Artificial intelligence on opportunistic computed tomography for predicting vertebral fracture risk in women undergoing estrogen deprivation therapy for breast cancer

Abstract

Purpose

Fracture risk prediction in women exposed to hormone deprivation therapies (HDTs) for breast cancer is challenging, since bone mineral density may not be reliable in this context. This study aims to identify radiomic features (RFs) on opportunistic computed tomography (CT) images associated with vertebral fractures (VFs) in women under HDTs and to develop a radiomics-based model predictive of VFs.

Methods

Radiomics analyses were performed on CT scans of 109 women (median age 61.1 years, range 27–85) exposed to HDTs (median duration of therapy 27.1 months). Lumbar vertebrae were automatically segmented (convolutional neural network) for RFs’ extraction. Each feature was tested for its ability to predict VFs, as assessed by a longitudinal morphometric approach before and during HDTs. Patients were randomly divided into training and test cohorts for the development and validation of the predictive model.

Results

During HDTs, new VFs were diagnosed in 23 women (21.1%) in association with older age (P = 0.013), lower total hip T-score (P = 0.041), and higher FRAX score for major fractures (P = 0.045). The machine learning (ML) model based on 20 RFs showed a high ability to predict VFs (ROC 0.832), outperforming that of T-score and FRAX score, even when thresholds lower than conventional ones were used (ROC 0.77 and 0.45, respectively). The RF "information measure of correlation" was the most relevant feature in the model, suggesting that a reduction in texture cross-correlation is positively associated with the development of VFs (P < 0.001).

Conclusion

The radiomics-based ML model showed high potential in identifying women at high fracture risk during HDTs.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11547-026-02172-6.

Introduction

Breast cancer (BC) is the most common cancer in women worldwide [1]. Approximately 80% of BC is hormone receptor positive and may be amenable to hormone deprivation therapies (HDTs), which are usually administered early in the course of the disease [2]. Current standard adjuvant HDTs for BC include tamoxifen (TAM), aromatase inhibitors (AIs), and gonadotropin-releasing hormone agonists (GnRHa), administered as monotherapy or in combination [3]. These therapies work by reducing the production or activity of serum estrogens, thereby reducing the main stimulus for cell proliferation in breast tissue, with beneficial effects in preventing cancer recurrence [4]. The depletion of circulating estrogen levels by HDTs has side effects on bone health, including a marked acceleration of bone resorption, a decrease in bone mineral density (BMD), and ultimately an increased risk of fracture [5]. Moreover, in addition to established risk factors for bone fragility, patients with BC have specific risk factors for fragility fractures, including increased adiposity [6], the direct negative effects of chemotherapeutic drugs and radiotherapy on osteoblasts and osteoclasts, concomitant medications such as high doses of glucocorticoids, and fatigue-related immobility [7, 8].

Recent joint position statements and updated international guidelines agree that the assessment of fragility fracture risk factors, biochemical testing for bone turnover, and baseline assessment of BMD should be performed in all women initiating HDTs [9–11]. The WHO Fracture Risk Assessment Tool (FRAX) and dual-energy X-ray absorptiometry (DXA) measurement of BMD have been proposed for identifying cancer patients at higher risk of fractures to be treated in primary prevention with bone-active drugs [11]. Specifically, the FRAX algorithm was developed to predict an individual’s 10-year probability of major osteoporotic fracture and hip fracture from readily assessed clinical risk factors and optional BMD [12]. However, FRAX and BMD have been validated for the assessment of fracture risk in postmenopausal osteoporosis, and it is still debated whether they could be reliable in the clinical setting of bone fragility due to HDTs [9]. Specifically, BMD assessment appears to be imprecise in predicting fragility fractures in this context because HDTs not only reduce bone mass but also induce early changes in bone quality that cannot be detected by DXA measurements of BMD [13]. Moreover, the FRAX algorithm does not include all risk factors for fractures in women exposed to HDTs, such as high body mass index (BMI) and impaired bone structure at the lumbar spine [13, 14]. As a result, fragility fractures can occur even in women with normal BMD and a FRAX score for major fractures below the conventional threshold of 20% [15–17]. Consequently, lower thresholds for both tools have been proposed for women undergoing HDT, though their accuracy in predicting fracture risk remains limited [9, 18].

The radiomics approach—defined as the high-throughput extraction of a large number of image features—has recently garnered significant interest for its ability to reveal disease characteristics that are challenging to capture with conventional tools [19, 20]. Emerging evidence suggests that radiomics performed on computed tomography (CT) scans could provide accurate information about the texture and spatial heterogeneity of bone microarchitecture, providing a comprehensive assessment of both bone quality and quantity [21–23]. Preliminary studies have found a potential application of a radiomics-based model in predicting fracture risk in postmenopausal osteoporosis [24]. In the specific setting of bone fragility in women with BC under HDTs, a recent study proposed novel CT radiomics-based opportunistic screening methods to detect BMD changes in BC patients, while no data are available on the possible correlation between radiomic features and fragility fractures [25].

This manuscript explores the application of radiomics and artificial intelligence (AI) algorithms to opportunistic CT imaging for predicting vertebral fracture risk in women undergoing HDTs for BC. Our aim is to evaluate the feasibility and accuracy of AI-enhanced CT analysis in detecting early changes in vertebral integrity that could predict fracture risk in this high-risk patient group. We hypothesize that radiomics and AI can provide valuable insights to guide clinical decision-making, leading to better outcomes through early intervention and personalized care. Specifically, the present study aims to identify radiomic features on opportunistic CT scans associated with vertebral fractures (VFs) in women undergoing HDTs and to develop a radiomics-based AI model to predict VFs.

Materials and methods

Study population

In this retrospective, single-center study, we performed an AI analysis based on radiomics of opportunistic CT scans of the lumbar spine in 109 consecutive women with breast cancer followed at the Metabolic Bone Diseases and Osteoporosis Section of IRCCS Humanitas Research Hospital between March 2018 and November 2023.

The inclusion criteria were: (1) age > 18 years; (2) histological diagnosis of BC; (3) subject undergoing HDTs; (4) availability of CT of the thorax and abdomen performed in our Institution for disease follow-up or for other reasons at least 6 months after HDTs initiation; (5) at least 2 outpatient visits during a period of at least 18 months after starting HDTs; (6) availability of at least two spine sagittal images (either DXA or X-rays or CT) to perform vertebral morphometry longitudinally during the study period. Exclusion criteria were: (1) bone metastases; (2) treatment with bone-active drugs (except calcium and vitamin D) at the time of the CT scan used for radiomics analysis.

The primary objective was to identify CT-based radiomics features able to predict VFs in women with BC exposed to HDTs and to implement an algorithm (scoring system) associated with the highest risk of developing VFs, as defined by a machine learning framework based on radiomics CT features. As secondary endpoints, we evaluated 1) baseline clinical parameters associated with the development of VFs; 2) a machine learning model predicting VFs based on clinical parameters; 3) the predictive value of the radiomics-based machine learning model compared to that of DXA and FRAX score. The complete workflow of the study is reported in Fig. 1.

Fig. 1.

Study design

All procedures in the present study that involved human participants were performed in accordance with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. The study was approved by the Ethics Committee of IRCCS Humanitas Research Hospital, and the patients gave their informed consent to use the clinical data for research purposes (Protocol code: NEURAD-2023-001; Project Code: RF-2021-12374134).

Clinical information

The following information was retrospectively collected from the clinical records: BC therapies and type of HDTs, family history of hip or VFs, personal history of clinical fractures, smoking (≥ 10 cigarettes/day), alcohol consumption (≥ 30 g/day), BMI, non-vertebral fractures during the study period, calcium and vitamin D supplementation, and treatment with potentially osteoporosis-inducing drugs. Assessment of fracture risk by FRAX algorithm (www.sheffield.ac.uk/FRAX) was performed in all enrolled patients with clinically adequate data at the time of the first clinical assessment. The input into FRAX included, with age and sex, BMD independent clinical risk such as BMI, previous fragility fracture (including morphometric vertebral fracture), parental history of hip fracture, current glucocorticoid treatment (any dose, by mouth for 3 months or more), current smoking, alcohol consumption, rheumatoid arthritis, secondary causes of osteoporosis [12]. The calculation of FRAX score was performed including BMD values of subjects and considering HDTs as a cause of secondary osteoporosis. Subjects with major osteoporotic fracture risk ≥ 20% were defined as high risk [12]. The exploratory analyses were also performed using the cutoff FRAX score for major fractures of 6.35%, which was already calculated as the best threshold for predicting fractures in the context of HDTs [18].

Assessment of VFs

Vertebral fractures were longitudinally detected twice during the study period by a quantitative morphometric assessment: (1) at baseline (at least 6 months from the start of HDTs) using DXA images and/or conventional spine X-ray radiographs and/or opportunistic spine computed tomography (CT) scans and (2) at follow-up, 18 months after baseline, using follow-up CT scans and/or DXA and/or X-rays. For morphometric VF assessment, anterior (Ha), middle (Hm), and posterior (Hp) vertebral heights were measured, and height ratios (Ha/Hp, Ha/Hm, and Hm/Hp) were calculated for each vertebra from T4 to L5 [26]. The evaluation of VFs was performed by an experienced physician (G.M.) who was blinded to the results of radiomics analyses. Positive findings were independently reviewed by a second evaluator (M.F.B.), and discordant cases were resolved by consensus.

Measurement of BMD-DXA examination

All enrolled patients were evaluated by DXA measurement of BMD at the lumbar spine, femoral neck, and total hip. BMD was expressed as T-score, comparing the results with those obtained in a gender-matched Caucasian population at the peak of bone mass. According to the WHO guidelines, a T-score ≤ − 2.5 SD at any skeletal site was defined as osteoporosis, whereas osteopenia was defined as a T-score between − 1 and − 2.5 SD [27].

CT acquisition protocol

Noncontrast 3D-CT scans were employed for RFs’ extraction in all enrolled subjects. These images were obtained through three different CT scanners: GE Revolution CT [acquisition parameters: KVp 120, mA adaptive] equipped with 256 rows of detectors, GE Evolution CT [acquisition parameters: KVp 120, mA adaptive], 64 rows of detectors, and Philips Ingenuity [acquisition parameters: KVp 140, mA adaptive], equipped with 64 rows of detectors. The abdomen Kernel standard reconstruction algorithm was employed for feature extraction and evaluation.

Image processing and radiomics feature extraction

Radiomics features were extracted from CT scan images acquired at baseline (at least 6 months from the start of HDTs). Non-fractured lumbar vertebrae (L1–L5) were automatically segmented from the non-contrast-enhanced CT images through a deep learning convolutional neural network (CNN) approach that allows both segmenting the vertebrae and identifying the vertebral level (Fig. 2). An in-house nn-UNet neural network model, already in use in the Neuroradiology Department of IRCCS Humanitas Research Hospital, was employed [23]. This segmentation method allows radiomics analyses of the whole vertebrae (including the vertebral body, lamina, spinous and transverse processes, pedicles, etc.), thus increasing the possibility of depicting modifications of the bone structure. The image segmentation was then validated by an experienced radiologist (L.S.P.).

Fig. 2.

Results of the automatic segmentation deep learning CNN of lumbar vertebrae (L1–L5 with different colors) in axial (A), coronal (B), and sagittal (C) views

A total of 107 radiomics features were extracted from each lumbar vertebra using the pyradiomics library (version 3.1.0), in compliance with the Image Biomarker Standardization Initiative (IBSI) [28].

Of the 107 extracted features, 18 were classified as the first-order statistics features, representing the statistical distribution of voxel intensities within the region of interest. The remaining 89 were classified as texture features (or second-order features, n = 75)—evaluating the spatial arrangement of voxel intensities and their relationships in the region of interest—and shape-based features (n = 14)—descriptors of the 2D or 3D size of the region of interest.

Each of the 107 features was reported both at the individual vertebral level (n = 535, resulting from the product of 107 features and 5 vertebrae) and at the aggregated level (n = 107), where the mean value across lumbar vertebrae was used, excluding any fractured vertebrae. Specifically, for patients without vertebral fractures (VFs), the mean was computed across all five lumbar vertebrae (L1–L5), whereas for patients with VFs, the fractured vertebrae were excluded from the aggregation. This aggregation strategy was adopted to assess whether radiomics features could capture not only localized changes but also diffuse bone alterations associated with treatment-related degeneration. In particular, features may reflect generalized skeletal deterioration when averaged across multiple vertebrae, providing complementary information to single-vertebra analysis.

In total, by considering both the individual vertebral features (n = 535) and the aggregated features (n = 107), we obtained 642 feature values for comparison between patients with and without VFs. In order to include also vertebral level in patients with VFs, an iterative imputer strategy based on k-NN algorithm (k = 8) was computed on the training set and applied on the test set. The algorithm was chosen to identify the ‘k’ most similar patients in the dataset for whom the L5 feature value is available.

Radiomics and clinical features selection for machine learning model

The CNN for vertebra segmentation was trained after assessing the agreement between two experienced radiologists’ manual segmentations, using the Dice coefficient to ensure high consistency in the reference annotations. The most discriminative radiomics and clinical features (i.e., those that have an impact on the accuracy of the prediction of vertebral fractures) were selected by applying bootstrap recursive feature elimination (RFE) procedures with random sampling for the train split (100 iterations). Bootstrapping has been performed between age-matched patients with and without VFs. At each iteration, the most important features had been saved, and after 100 iterations, 15 radiomics features and 3 clinical features with the highest occurrence had been selected.

Machine learning model

For the development of the machine learning models for VFs’ prediction, patients were split into training cohort (60%) and test cohort (40%), stratified by VFs’ outcome. The Tree-Parzen estimator Bayesian approach was used, and a cross-validated feature selection algorithm was employed on the training cohort to ensure that features’ importance was independent on the specific subset of training bootstrap group. The RF-based model was obtained by training a logistic regression model, while the clinical-based model was trained on a random forest classifier. We also computed two combined RF and clinical-based models: one model aggregating at features’ level (early fusion) and the other evaluating a weighted classifier by averaging probabilities from RF and clinical-based models. The four machine learning models evaluated on the test cohort and accuracy, sensitivity, specificity, and area under the ROC curve were obtained. The final predicted probability was then converted for better clinical usability into a 0–10 point score by multiplying by 10 (i.e., score = 10 × Probability).

Statistical analysis

Data were presented as mean ± standard deviation, unless otherwise stated. The Shapiro test was preliminarily employed to assess the normality distribution of the data. To compare the case (i.e., patients with VFs) with the control group (i.e., patients without VFs), univariate statistical analysis was performed using a two-tailed Student’s t test on normally distributed clinical and radiomic features, whereas the two-tailed Mann–Whitney test was applied to non-normally distributed features.

Inter-rater reliability of the extracted radiomics features was quantified using the intraclass correlation coefficient (ICC), calculated by comparing feature values derived from CNN-based segmentations and those from manual segmentations performed by experienced radiologists. P value < 0.05 was considered statistically significant for the statistical tests used in our analysis. All analyses were made with Python (version 3.11.0).

Results

Cohorts’ presentation

One-hundred and nine women (mean age 61.1 years, SD ± 11.3, range 27–85) were included in the study. At the time of the first clinical assessment, 20 women (18.3%) were under treatment with GnRHa plus AI, 6 (5.5%) on GnRHa plus tamoxifen, 2 (1.8%) on GnRHa alone, 76 (69.7%) on AI alone, and 3 (2.8%) on tamoxifen alone, and in 2 cases (1.8%), information on HDTs’ type was not recorded. The mean duration of HDTs’ exposure before performing CT scans used for radiomic analysis and to assess the occurrence of new VFs was 27.1 months (range 6–63). Before HDTs’ initiation, 51 (46.8%) and 80 (73.4%) women underwent chemotherapy and radiotherapy for BC, respectively. Six women (5.5%) were under glucocorticoids therapy (4 with topic corticosteroids for asthma and 2 with systemic corticosteroids for coexistent rheumatoid arthritis). At baseline assessment, 105 (96.3%) and 55 (50.5%) women were under vitamin D and calcium supplementation, respectively.

At the first visit, BMD assessment was available in 100 out of 109 enrolled women (91.7%); 34 of them (34%) had osteoporosis at either skeletal site (22 women with osteoporosis only at lumbar spine, 12 only at the hip, and 6 at both skeletal sites), whereas osteopenia and normal BMD at all skeletal sites were found in the 45 (45%) and 21 (21%) women, respectively. At the first visit, FRAX score was calculated in 99 out of 109 enrolled women (90.8%); 14 women (14.1%) had a FRAX score for both 10-year risk of major fractures ≥ 20%, while the remaining 85 women (85.8%) had the FRAX score under the currently used cutoff of 20%. New VFs were diagnosed in 23 women (21.1%). Women with VFs were older (66 ± 10.0 vs. 59.7 ± 11.3, P = 0.013) and had lower total hip T-score value (− 1.5 ± 1.0 vs. − 0.8 ± 1.0, P = 0.041) and higher FRAX score for major fractures (14.5 ± 9.92 vs. 10.0 ± 7.3, P = 0.045) as compared to those without VFs, without significant differences in previous chemotherapy and radiotherapy treatment (P = 0.611 and P = 0.563, respectively), concomitant glucocorticoid therapy (P = 1.00), positive family history for hip or vertebral fractures (P = 0.861), history of premature menopause (P = 0.176), diabetes mellitus (P = 1.00), cigarette smoking (P = 0.691), and alcohol consumption (P = 0.243) (Table 1).

Table 1.

Baseline clinical and demographical data of 109 premenopausal women with BC under HDTs, stratified according to the occurrence of vertebral fractures

	Overall	No VFs	VFs	P value
Cases (N)	109	86	23
Age (years), mean (SD)	61.1 (11.3)	59.7 (11.3)	66.0 (10.0)	0.013
HDTs exposure (months), mean (SD)	27.1 (15.5)	26.9 (14.8)	27.9 (18.1)	0.812
Type of HDTs (N, %)				0.325
AIs	76 (71.0)	57 (67.1)	19 (86.4)
GnRHa	2 (1.9)	2 (2.4)
GnRHa + AIs	20 (18.7)	18 (21.2)	2 (9.1)
GnRHa + Tamoxifen	6 (5.6)	6 (7.1)
Tamoxifen	3 (2.8)	2 (2.4)	1 (4.5)
Chemotherapy (N, %)	51 (48.6)	42 (50.0)	9 (42.9)	0.611
Radiotherapy (N, %)	80 (74.8)	62 (72.9)	18 (81.8)	0.563
Family history of hip Fx/VFs (N, %)	28 (28.9)	22 (27.8)	6 (33.3)	0.861
Premature menopause (N, %)	26 (40.6)	24 (45.3)	2 (18.2)	0.176
Cigarettes smoking > 10/day (N, %)	11 (10.7)	8 (9.8)	3 (14.3)	0.691
Diabetes mellitus type 2 (N, %)	6 (5.7)	5 (6.0)	1 (4.5)	1.000
Glucocorticoid therapy (N, %)	6 (7.1)	5 (7.1)	1 (6.7)	1.000
BMI (kg/m2), mean (SD)	25.3 (4.7)	25.2 (4.7)	25.6 (5.0)	0.761
Alcohol consumption > 3 U/day (N, %)	2 (3.2)	1 (1.9)	1 (12.5)	0.243
Vitamin D supplementation (N, %)	105 (97.2)	82 (96.5)	23 (100.0)	1.000
Calcium supplementation (N, %)	55 (50.9)	40 (47.1)	15 (65.2)	0.190
L1–L4 T-Score, mean (SD)	− 1.5 (1.4)	− 1.5 (1.4)	− 1.5 (1.6)	0.980
T-Score FN, mean (SD)	− 1.4 (0.9)	− 1.4 (0.9)	− 1.4 (1.0)	0.913
T-Score TH, mean (SD)	− 0.9 (1.1)	− 0.8 (1.0)	− 1.5 (1.0)	0.041
DXA diagnosis (N, %)				0.825
Normal BMD		16 (19%)	5 (22%)
Osteopenia		37 (43%)	8 (35%)
Osteoporosis		25 (29%)	7 (30%)
FRAX score major fracture, mean (SD)	10.9 (7.9)	10.0 (7.3)	14.5 (9.1)	0.045
FRAX score hip fracture, mean (SD)	2.8 (4.7)	2.4 (4.2)	4.4 (6.0)	0.162

Continuous unpaired data were presented as mean ± standard deviation, whereas categorical data were presented as number of cases and percentages. The values in bold refer to those statistically significant. Abbreviations: AIs, aromatase inhibitors; BMI, body mass index; Fx: fractures, GnRHa, gonadotropin-releasing hormone agonists; HDTs, hormone deprivation therapies; N, number; VFs, vertebral fractures

To build and evaluate a radiomics-based machine learning model predictive of VFs, patients were randomly assigned either to a training (n = 65) or to a test (n = 44) cohort. No relevant clinical differences were detected between training and test cohorts (Supplementary Table S1).

Radiomics and clinical features and correlation with VFs

In the univariate analysis, 90 out of 642 feature values calculated from lumbar unfractured vertebrae were significantly associated with VFs. Of these, 69 were feature values calculated at the single vertebrae level and 21 were feature values calculated at the aggregated level. Focusing on the aggregated level, significant differences were found between women with and without VFs in terms of first, second, and shape-based features (Supplementary Table S2). In detail, women with at least one VF had lower first-order energy as compared to those without a VF (P = 0.046). Regarding textural features, patients with VFs had higher gray-level co-occurrence matrix (GLCM) correlation, informal measure of correlation 1 and 2, inverse difference normalized, maximal correlation coefficient, inverse different moment normalized (P < 0.01 for all comparisons), and inverse variance (P = 0.049), lower GLCM difference entropy (P = 0.043), lower neighboring gray-tone difference matrix (NGTDM) busyness and coarseness (P = 0.015 and P = 0.016, respectively), higher NGTDM strength (P < 0.05), higher gray-level run length matrix (GLRLM) run variance (P = 0.035), lower gray-level size zone matrix (GLSZM) size zone non-uniformity normalized, zone percentage, and small area emphasis (P < 0.05 for all comparisons) as compared to patients without VFs. For shape-based features, women with VFs had lower maximum 2D diameter row and major axis length and higher flatness (P < 0.05 for all comparisons). Twenty radiomic features of single vertebrae were found to be the most contributive for the development of the machine learning model. Specifically, the top first-order, second-order, and shape-based features with the highest weight for the logistic regression model are highlighted in Fig. 3A. The same approach was used to select clinical features predictive of VFs, and age, calcium supplementation, and total hip T-score resulted as those with the highest weight for the random forest model (Fig. 3B).

Fig. 3.

The radiomic (A) and clinical (B) features that contributed to developing the machine learning model

Performance of machine learning model

The performance of the four developed machine learning models in identifying patients with VFs is shown in Fig. 4. As shown, the ability of the radiomics-based model to predict VFs was higher than that of the model based on clinical features (radiomics-based model training ROC 0.881, test ROC 0.832, clinical-based model training ROC 0.619, test ROC 0.733). Notably, the predictive ability of the machine learning model did not improve when both clinical and radiomic features were integrated in early fusion or weighted-classifier strategies (Fig. 4).

Fig. 4.

The ROC analysis assessing the performance of the machine learning model on the test cohort in predicting VFs, based either on radiomics or clinical features

We selected 5.5 as the optimal cutoff of the final model by analyzing accuracies obtained on the training set. Using this cutoff, the final model obtained good accuracies in identifying patients at high risk of fracture in the test cohort (accuracy 0.75, sensitivity 0.77, specificity 0.74). Using the cutoff of 5.5, only 2 patients of the test cohort who experienced VFs had a radiomics-based score below this threshold. Notably, these two fractured patients, who were classified as "low fracture risk" by the radiomics-based model, had BMD values indicative of osteoporosis. With respect to CT acquisition protocol parameters, no significant differences in the model’s predictive performance were observed between patients scanned at 120 kV (accuracy = 0.76, sensitivity = 0.82, specificity = 0.73, n = 37) and those scanned at 140 kV (accuracy = 0.71, sensitivity = 0.75, specificity = 0.67, n = 7).

Radiomics-based model versus currently used tools

We then compared the performance of the radiomics-based model in predicting VFs to that of tools currently used in clinical practice, namely T-score BMD and FRAX score. We used total hip T-score and FRAX score for major fractures as they resulted significantly associated with VFs in our whole cohort. As shown in Fig. 5, the ability of the radiomics-based model in detecting fragility VFs resulted higher as compared to that of T-score and FRAX score (ROC 0.665 and ROC 0.635 in the test cohort, respectively). Specifically, in our test cohort, a T-score total hip of –2.5 showed limited ability in VFs (accuracy 0.54, sensitivity 0.89, specificity 0.45). Moreover, also applying the lower cutoff of − 1—proposed by guidelines on the management of bone health in women with BC on HDTs—the predictive ability remained low (accuracy 0.77, sensitivity 0.07, specificity 0.97). Regarding FRAX score for major fractures, the ability in predicting VFs remained limited both by applying the standard cutoff of 20% (accuracy 0.77, sensitivity 0.33, specificity 0.88) and the recently proposed lower cutoff for women on HDTs of 6.35% (accuracy 0.45, sensitivity 1, specificity 0.31).

Fig. 5.

The ROC analysis assessing the performance in predicting VFs on the test cohort of radiomics-based machine learning model, FRAX score for major fractures and total hip T-score

Discussion

This retrospective study showed that a machine learning model based on radiomics analyses computed on opportunistic lumbar CT scans has great potential for predicting skeletal fragility in women with BC undergoing HDTs. The predictive model based on first-order, second-order, and shape-based radiomics features accurately identified patients who experienced VFs during follow-up, outperforming the machine learning model based on clinical features and BMD and FRAX scores, currently used for fracture risk assessment.

Women with BC on HDTs are at high risk of skeletal fragility, but it is still debated how to properly assess fracture risk in this setting [9]. Recent data have clearly demonstrated that BMD assessment may not be reliable in predicting fragility fractures in this clinical context as many fractures occur in patients with BMD values in the osteopenia or even normal reference range [15, 17]. In our cohort, approximately 20% of women exposed to HDTs developed morphometric VFs, an incidence consistent with previous literature data [14, 15, 29, 30]. Although patients with VFs had a lower total hip T-score than those who did not develop VFs, the mean value was in the osteopenia range, confirming the recent suggestion that a T-score threshold of − 1.0 may be more reliable for predicting fracture risk in patients on HDTs than the conventional − 2.5 [9, 18]. Of note, although the femoral neck T-score is more commonly used in clinical practice, only total hip T-score was associated with VFs and selected by bootstrap recursive feature elimination (RFE) for our machine learning model based on clinical features. This is consistent with meta-regression analyses in osteoporotic women and men, reporting that total hip BMD measurement is strongly associated with fracture risk, more reproducible than that of the femoral neck and more reliable than that of the lumbar spine [31]. For the machine learning model based on clinical parameters, older age, and lower rate of calcium supplementation, both well-recognized determinants of skeletal health [32, 33] were selected as the most contributing features in addition to total hip T-score. Although patients with VFs also had higher FRAX scores for major fractures, the FRAX score did not fit in the predictive model. Notably, as with the T-score, the mean FRAX score for major fractures in patients with VFs was lower than the conventional threshold of 20% used to identify patients at high risk of fracture. This is consistent with recent data suggesting that the conventional threshold of 20% severely underestimates fracture risk in women on HDTs and that a lower threshold should be used [18]. In addition, the FRAX score does not appear to be the right tool for assessing risk factors in this specific setting, as it has not been validated in women younger than 40 years and was not designed for patients starting HDTs [9]. The case of BMI is emblematic in this regard, as it is considered a risk factor in the FRAX tool if it is below 20 kg/m², but in contrast to postmenopausal osteoporosis, high BMI has been shown to be associated with fracture risk in women on HDTs [6, 13, 16, 17].

The limited predictive value of these conventional tools and our machine learning model based on clinical features is explained by the evidence that HDTs cause early changes in bone microarchitecture with consequent deterioration in bone quality, which is independent of BMD loss and cannot be properly assessed by DXA assessment [13]. Otherwise, the machine learning model developed in the present study, based on first-order, second-order, and shape-based radiomics features, demonstrated a high ability to predict VFs, as these features reflect both bone quality and quantity characteristics [23]. Specifically, first-order features describe the distribution of individual voxel intensities within a volume of interest (VOI), and these features have been proposed as indirect surrogates of BMD in the specific VOI [34]. Second-order features, also known as texture features, provide information about the spatial relationship between voxels in the VOI and are considered markers of bone quality and trabecular microarchitecture [35–37]. Finally, shape-based features are quantitative descriptors of the 2D and 3D size and shape of the VOI and have recently been proposed as markers of geometric parameters, including trabecular thickness and bone volume fraction [38]. The majority of the radiomics features included in our machine learning model were second-order and shape-based, consistent with the hypothesis that HDTs affect bone quality earlier and more than bone quantity. In particular, features belonging to the gray-level co-occurrence matrix (GLCM) category were the most relevant to the predictive model [39]. These features quantify how often pairs of pixel intensities (gray levels) co-occur in a given VOI and are reminiscent of the Trabecular Bone Score (TBS), a texture index that reflects gray-level variations and is well recognized as a marker of bone quality [40]. In this respect, the reduction in texture cross-correlation found in our patients with VFs may correspond to the decrease in TBS found in both premenopausal and postmenopausal women [14, 41]. Thus, the radiomics approach, which appears to detect the profound bone changes induced by HDTs, outperformed the ability of both the clinical feature-based model and the BMD and FRAX scores in predicting VFs. Notably, the predictive value of the radiomics-based model remained superior to that of these tools even when applying the T-score of less than − 1.0 and the lower FRAX score of 6.35% recently proposed for the specific setting of women on HDTs [18]. Our results suggest that the implementation of this model in clinical practice could be useful in identifying women undergoing HDT who should be treated with bone-active drugs in primary prevention, especially those who are considered "at low risk" based on the assessment of BMD and FRAX scores. In the scoring system based on radiomic features, we identified 5.5 as the optimal threshold to identify patients at high risk of fracture, as it provides a better balance between sensitivity and specificity. Indeed, in our cohort, using this threshold, the two women falsely predicted to be at low risk of fracture had a densitometric diagnosis of osteoporosis and would have been treated anyway, as they were by definition at high risk.

Finally, it should be noted that the radiomic features of individual lumbar vertebrae were able to predict VFs in the entire T4–L5 tract. Thus, if further validated, the use of this model in clinical practice could be hypothesized independent of the availability of thoracoabdomical CT scan, as a narrow lumbosacral CT imaging acquisition would be sufficient to predict fracture risk, thus limiting radiation exposure.

Our study has several limitations, mainly related to its retrospective nature. The first inherent limitation of this single-center pilot study was the small sample size and the lack of an external validation cohort. Secondly, the wide range of exposure to HDTs prior to performing the CT scans used for radiomics analyses, which may have influenced the type and number of features associated with VFs. Thirdly, the lack of TBS and full biochemical assessment including bone turnover markers, which did not allow confirmation of the change in bone quality presumably represented by the second-order and shape-based features included in our model. Finally, CT acquisition parameters were not included as a correction factor during radiomic feature extraction due to the imbalance in the number of cases between the 120 kV and 140 kV protocols. However, our study also has several strengths, including the use of opportunistic CT scans, the rigorous methods used for segmentation and feature extraction and model validation, and the novelty of the approach. In this regard, only one recent study has been published proposing a radiomics-based model to predict cancer treatment-induced bone loss in women with BC. However, the radiomic features were investigated in relation to BMD parameters rather than fragility fractures [25]. To our knowledge, this is the first study to propose a radiomics-based machine learning model able to predict VFs in the specific setting of hormone deprivation-induced skeletal fragility.

Conclusions

The radiomics-based machine learning model developed in the present study showed high potential in predicting fragility fractures in women with breast cancer undergoing hormone deprivation therapy, outperforming the predictive value of currently used fracture risk assessment tools. A prospective approach in larger cohorts is advocated to validate and optimize our model.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Study conception and design were performed by Letterio Salvatore Politi, Gherardo Mazziotti, and Riccardo Barbieri. Data collection was performed by Maria Francesca Birtolo, Marta Pagano, Flaminia Carrone, Simone Antonini, and Giulia Maida. Material preparation and analysis were performed by Riccardo Levi, Giovanni Savini, Francesca Geroli, and Simona Superbi. The first draft of the manuscript was written by Maria Francesca Birtolo, Riccardo Levi, Gherardo Mazziotti, Andrea Lania, and Letterio Salvatore Politi, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the Italian Ministry of Health, RF-2021-12374134.

Declarations

Competing interests

The authors have no relevant financial or non-financial interests related to the submitted work to disclose.

Research involving human participants and/or animals

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of IRCCS Humanitas Research Hospital (Protocol code: NEURAD-2023-001; Project Code: RF-2021-12374134).

Informed consent

All the patients gave their informed consent to use the clinical data for research purposes and publication.

Ethics standards

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of IRCCS Humanitas Research Hospital, and the patients gave their informed consent to use the clinical data for research purposes (Protocol code: NEURAD-2023-001; Project Code: RF-2021-12374134).