In order to distinguish enchondromas from ACTs using MRI, this study sought to investigate the capabilities of combining radiomics with machine learning techniques. Our findings suggest that the analysis of quantitative radiomics parameters using machine learning algorithms offers a promising potential to improve diagnostic accuracy, even though it can be challenging to distinguish between these cartilage tumours using conventional imaging alone. Remarkably, the approach we employed produced AUC values higher than 0.90, with the XGBoost model obtaining a noteworthy AUC of 0.987. This robust performance underscores the potential of our non-invasive method to offer valuable insights in situations where traditional assessment may be ambiguous, thereby assisting clinicians in making more informed decisions about patient management.

The distinction between enchondroma and ACT poses a significant challenge for radiologists, pathologists, and clinicians in clinical practice, despite numerous attempts in the literature to overcome this difficulty [26]. Conventional radiological methods have shown limited success in this differentiation, primarily due to high rates of both false-positive and false-negative diagnoses [7]. A 2012 study by Ferrer-Santacreu et al. investigating the role of MRI in differentiating the two entities in the appendicular skeleton found no statistically significant distinguishing MRI features [27]. These findings were contradicted a year later by Choi et al., who identified several MRI features suggestive of malignancy, including a predominantly intermediate signal intensity matrix on T1-weighted imaging, a multilobulated enhancement pattern, cortical destruction, an associated soft tissue component, surrounding bone marrow and/or soft tissue oedema, epiphyseal involvement, and/or location in a flat bone [10]. In 2015, Crim et al. assessed the efficacy of MRI and radiography in distinguishing enchondroma from low-grade chondrosarcoma in 53 patients. In assessing low-grade lesions, they discovered that both radiographs and MRI had limitations, and bone expansion and cortical thickening were uncommon but distinct signs of malignancy [7]. Moreover, it has been reported that there is considerable overlap in the enhancement patterns of enchondromas and low-grade chondrosarcomas on dynamic contrast-enhanced MRI, limiting its value in differentiation [9]. This highlights the controversy in the literature, as other studies, such as the one by De Coninck et al., have suggested it to be a useful tool [28]. Discrimination by conventional strategies continues to be considered a diagnostic challenge. As of right now, no documented standardized management guidelines have been published.

Recent advancements in radiomics and machine learning (ML) offer promising avenues for differentiating enchondroma from chondrosarcoma. Gitto et al. (2021) investigated the influence of interobserver variability on 2D and 3D CT- and MRI-based texture feature reproducibility specifically in cartilaginous bone tumours, highlighting that radiomics features extracted from unenhanced CT and MRI are reproducible, despite some interobserver segmentation variability [29]. Pan et al. (2021) further contributed by developing radiomics nomograms based on non-enhanced MRI and clinical risk factors (tumour location, age, sex), achieving high AUCs (up to 0.961 in the validation group) for differentiating chondrosarcoma from enchondroma, particularly emphasizing the strong performance of T1-weighted and combined T1-weighted + T2-weighted features [30]. Erdem et al. (2023) demonstrated that neural networks, among seven ML models, achieved high accuracy (AUC 0.979–0.990) in distinguishing these lesions using radiomics features from T1 and PD-weighted MRI [31]. Cilengir et al. (2023) reported promising diagnostic value for MRI-based texture analysis, with the K-neighbours classifier achieving an AUC of 1.00 for FS-PD images in differentiating enchondroma and chondrosarcoma [32]. Although the systematic review conducted by Zhong et al. (2023) concluded that the quality of radiomics studies in chondrosarcoma was insufficient (median Radiomics Quality Score: 10.5), it also noticed the potential for optimising clinical decision-making [33].

To our knowledge, this study is the first to specifically target the differentiation of enchondroma and ACT using radiomics and machine learning on MRI examinations. Although literature on radiomics research aimed at distinguishing enchondroma from low-grade chondrosarcoma is limited, different modalities aside from MRI have been studied. For instance, Yoon et al. investigated the value of SPECT/CT radiomics parameters in differentiating enchondroma and grade I chondrosarcomas in long bones [34]. Their study, involving 49 patients, found that a specific radiomics parameter (zone-length non-uniformity) was a significant independent factor for distinction, achieving a sensitivity of 83.3% and specificity of 90.9% in their test data. Nevertheless, their study was limited by the absence of machine learning algorithms and a relatively small sample size. Similarly, Yildirim et al. developed a CT radiomics-based machine learning model to differentiate low-grade chondrosarcoma from enchondroma [35]. This study, with 30 enchondroma and 26 chondrosarcoma patients, utilized 3D CT-based radiomics and various machine learning models. They reported that a random forest model using 5 parameters achieved an AUC of 0.967. A limitation of this study was also its retrospective nature, as well as the relatively small cohort size.

Our study’s focus on T1-weighted MRI further highlights its potential for widespread clinical applicability, given the common use and consistency of this sequence across various anatomical locations in musculoskeletal imaging. Radiomics approach employed in our study extracted 107 quantitative parameters from T1-weighted MRI images, which were subsequently reduced to 10 most discriminative parameters using LASSO regression. This dimensionality reduction technique helped identify the most relevant features while avoiding overfitting, a common challenge in radiomics studies. XGBoost (AUC = 0.987) performed better than other machine learning methods like SVM (0.922), Decision Tree (0.949), and even Random Forest (RFC: 0.92). This is most likely attributable to its gradient boosting approach. XGBoost builds a strong prediction model by combining many simple models, usually decision trees, one after another. This step-by-step improvement helps XGBoost find complex patterns in detailed medical image data, often doing better than single models or other ensemble methods like Random Forest. Its strength comes from L1 and L2 regularization techniques that prevent overfitting and make it more generalizable, which is key for medical data with limited samples [23]. The high specificity (96.55%) achieved by XGBoost is particularly noteworthy, as it suggests a low false-positive rate, which is crucial in clinical practice to avoid unnecessary surgical interventions for enchondromas.

The performance of our T1-based radiomics model may be explained by its ability to quantify subtle, pixel-level textural variations that are often imperceptible to the human eye, yet appear to correspond to key histopathological differences. ACTs are defined by features such as increased cellularity and a permeative growth pattern, which can disrupt the uniform signal of surrounding marrow fat and create a more complex, heterogeneous texture on T1-weighted images. An analysis of the ten most discriminative features selected by LASSO regression, detailed in Table 2, lends support to this hypothesis. For instance, a higher DependenceNonuniformityNormalized (coefficient: 0.180) was a strong predictor for ACT, suggesting that textural heterogeneity is a key indicator of malignancy. This is further supported by the Coarseness parameter (coefficient: -0.133); a lower coarseness value, which points to a finer and less uniform texture, was also associated with ACT. This could reflect the microscopic architectural disruption caused by permeative growth. Conversely, a high ShortRunLowGrayLevelEmphasis (coefficient: -0.326) was the strongest predictor for enchondroma, potentially reflecting its more organized and less aggressive structure. In essence, our model seems to move beyond conventional radiological assessment by translating these microscopic, visually elusive hallmarks of malignancy into a quantitative signature.

Our research possesses a few limitations. First, the study was a single-centre retrospective research with a relatively limited sample size. Multi-centre prospective studies with larger samples are needed to validate our findings and evaluate the generalizability of our radiomics models. Second, our analysis was limited to T1-weighted images and 2D segmentations. The inclusion of other MRI sequences (such as T2-weighted, STIR, or contrast-enhanced sequences) with the acquisition of 3D volume of interests (VOIs) might further improve diagnostic capability. Additionally, the development of automated segmentation could further streamline the clinical implementation of radiomics-based diagnosis. We also acknowledge that we did not systematically analyze intratumoural mineralization, a potential confounder given its effect on MRI signal. However, the textural heterogeneity caused by mineralization may also represent a valuable source of data for our radiomics parameters. The high accuracy achieved by the radiomics-based ML algorithms suggests that these complex patterns, while challenging for visual interpretation, may contain valuable information for quantitative classification.

In conclusion, our study demonstrates that MRI-based radiomics combined with machine learning, particularly XGBoost, can achieve robust diagnostic performance in differentiating enchondroma from ACT. While this non-invasive approach has the potential to significantly improve diagnostic accuracy and clinical decision-making, its findings must be validated by prospective studies with larger sample sizes.