Skip to main content
Download PDF

A deep learning model for classification of chondroid tumors on CT images

BMC Cancer volume 25, Article number: 561 (2025) Cite this article

Abstract

Background

Differentiating chondroid tumors is crucial for proper patient management. This study aimed to develop a deep learning model (DLM) for classifying enchondromas, atypical cartilaginous tumors (ACT), and high-grade chondrosarcomas using CT images.

Methods

This retrospective study analyzed chondroid tumors from two independent cohorts. Tumors were segmented on CT images. A 2D convolutional neural network was developed and tested using split-sample and geographical validation. Four radiologists blinded to patient data and the DLM results with various levels of experience performed readings of the external test dataset for comparison. Performance metrics included accuracy, sensitivity, specificity, and area under the curve (AUC).

Results

CTs from 344 patients (175 women; age = 50.3 ± 14.3 years;) with diagnosed enchondroma (n = 124), ACT (n = 92) or high-grade chondrosarcoma (n = 128) were analyzed. The DLM demonstrated comparable performance to radiologists (p > 0.05), achieving an AUC of 0.88 for distinguishing enchondromas from chondrosarcomas and 0.82 for differentiating enchondromas from ACTs. The DLM and musculoskeletal expert showed similar performance in differentiating ACTs from high-grade chondrosarcomas (p = 0.26), with an AUC of 0.64 and 0.56, respectively.

Conclusions

The DLM reliably differentiates benign from malignant cartilaginous tumors and is particularly useful for the differentiation between ACTs and Enchondromas, which is challenging based on CT images only. However, the differentiation between ACTs and high-grade chondrosarcomas remains difficult, reflecting known diagnostic challenges in radiology.

Peer Review reports

Background

Cartilaginous bone tumors, characterized by tumor cells that produce a chondroid matrix, are divided into benign enchondromas and malignant chondrosarcomas. Chondrosarcomas, the most common primary malignant bone tumors, are classified into various histological grades and subtypes [1, 2]. According to the World Health Organization classification system, Chondrosarcomas grade 1, located in the appendicular skeleton, are now termed “atypical cartilaginous tumors” (ACT) [3]. While enchondromas usually do not require any treatment, ACTs and chondrosarcomas necessitate curettage and/or wide resection [4]. Accurate classification of these tumors is therefore crucial for appropriate treatment.

The final diagnosis of cartilaginous tumors is based on a combination of clinical findings, histopathological results, and imaging studies, and is ideally performed at centers specialized in the treatment of musculoskeletal tumors [5, 6]. While biopsy remains essential in the diagnostic pathway of malignant cartilaginous tumors, the final diagnosis often relies on consensus reached through interdisciplinary tumor board discussions, integrating clinical history, histopathological findings, and radiological assessments [7, 8].

Several studies have investigated imaging parameters to improve the differentiation, especially between ACTs and enchondromas [9,10,11]. Murphey et al. showed that cortical destruction, periosteal reaction and endosteal scalloping (> 2/3 of cortical thickness) significantly increased the likelihood of an ACT compared to an enchondroma [12]. Nevertheless, radiological differentiation between enchondromas and ACTs remains challenging and highly variable depending on the radiologist’s expertise.

To improve diagnostic accuracy, a deep learning-based approach capable of precisely categorizing cartilaginous tumors as benign or malignant based on CT images only, may be useful for clinical routine. Recently, studies have shown that deep learning models (DLM) reliably assess and detect a variety of musculoskeletal diseases based on medical imaging data [13,14,15].

More recent studies have extended these findings by incorporating multimodal imaging approaches, such as combining X-ray, CT, and MRI, to enhance classification performance [16]. Additionally, radiomics-based machine learning models have demonstrated promising results in distinguishing enchondromas from atypical cartilaginous tumors (ACTs) and low-grade chondrosarcomas [17, 18].

Therefore, the aim of this study was to develop a DLM specifically designed to differentiate between enchondromas, ACTs, and high-grade chondrosarcomas based on CT images.

Methods

Approval of the Institutional Review Boards had been obtained prior to this study (Institutional Review Board of the Technical University of Munich, approval number 393/20, and the Ludwig-Maximilian University of Munich, approval number 21–0282). Written informed consent was waived for this retrospective analysis of routinely acquired imaging and clinical data. All analyses are in line with the declaration of Helsinki.

Data sets

In this study we evaluated cartilaginous tumors on CT images of patients from two university hospitals (Technical University of Munich and the Ludwig-Maximilian University of Munich) treated between 2011 and 2020. These patients had a final diagnosis of either an enchondroma, an ACT or a high-grade chondrosarcoma and preoperative CT images available in the respective picture archiving and communication systems (PACS). The diagnoses were determined based on the consensus of the respective local interdisciplinary tumor board consisting of specialized pathologists, radiologists, and orthopedic tumor surgeons. Patients with enchondroma, ACT or high-grade chondrosarcoma of the hand or foot were excluded from this study. The internal dataset from the Technical University of Munich included 244 patients, divided using split-sample validation into training/validation/hold-out testing sets of 60% (n = 146)/ 20% (n = 49)/ 20% (n = 49). The external test set using geographical validation was obtained from Ludwig-Maximilian University for further, independent testing with a total of 100 (29.07% of the entire dataset) patients. In total, 344 CT-imaging datasets were extracted of patients diagnosed with enchondromas (n = 124), ACTs (n = 92), and high-grade chondrosarcomas (n = 128).

For comparison of diagnostic performance between radiologists and DLM, two radiologic residents with 4 years of experience, one musculoskeletal fellow and one musculoskeletal attending radiologist with 10 and 30 years of experience, respectively, classified the tumors of the external dataset.

Portable Network Graphics-files were extracted from Digital Imaging and Communications in Medicine files for further image processing. For the training dataset, segmentations of the tumors were performed by L.K. and F.G.G. and reviewed by a musculoskeletal fellowship trained radiologist (A.S.G) using the open-source software 3D Slicer (version 4.7; www.slicer.org). The radiologists were blinded to the histopathological and clinical data. Segmentations were performed using the entire datasets and all slices on which the tumor was depicted. Figure 1 shows exemplary images of the segmentations of an enchondroma and an ACT. The image containing the largest tumor area was selected, followed by image cropping around the tumor based on the segmentation mask provided by expert radiologists. Thus, a 2D model was designed for image analysis.

Fig. 1
figure 1

Axial CT Images of patients with an enchondroma (A, B) and a high-grade chondrosarcoma (G2) (C, D) with (B, D) and without the respective segmentations (A, C)

Full size image

Model architecture and model training

A 2D convolutional neural network (CNN) was trained for classification. For hyperparameter selection, an empirical grid search has been performed, with each model being trained for 100 epochs. The final best model consisted of three convolutional layers of size 16, 32 and 64, each followed by a max pooling layer with a stride of 2 × 2, ending in a fully connected layer of size 128 followed by a single output neuron. Furthermore, an optimal learning rate of 0.001 and a batch size of 32 have been identified. Convolutional and pooling layers featured a kernel size of 3 × 3. Labels of 0.0, 0.5 and 1.0 were assigned to the three classes, reflecting the increase in severity. The mean squared error loss was used for optimization.

Statistical analysis

Calculations of model metrics and model evaluations were performed using Stata version 18.0. The performance of the models was evaluated with the area under the curve (AUC) obtained from the receiver operating characteristic (ROC) analysis on both the validation set and the independent hold-out test set. Furthermore, final model performance was evaluated using accuracy, sensitivity, specificity and AUC on the independent test-set as defined previously [19]. For performance comparison McNeil’s test was used, and statistical significance was defined as p < 0.05. Model training and evaluation was performed by D.L. (5 years’ experience in data analysis) and S.B. (12 years’ experience in data analysis). The intra- and interreader agreement of CT readings was assessed with Fleiss’ κ. To assess the intrareader reliability of the tumor segmentations, ten patients were randomly selected from the internal data set. One radiologist (L.K.) repeated the segmentations three months after the initial segmentation, blinded to the previous results. Segmentations were compared using the dice score.

Results

Patient characteristics and data sets

This study included 344 patients (mean age 50.3 ± 14.3 years; 175 women). There were no significant differences in age (p = 0.33) or sex (p = 0.36) between the internal and external datasets. Based on the consensus of the interdisciplinary tumor board, the internal dataset comprised 108 patients (44%) with enchondromas, 58 patients (24%) with ACTs, and 78 patients (32%) with high-grade chondrosarcomas. The external dataset included 16 patients (16%) with enchondromas, 35 patients (35%) with ACTs, and 49 patients (49%) with high-grade chondrosarcomas. Patient characteristics are summarized in Table 1.

In total, 101 tumors (29.4%) were located in the upper extremity, 130 tumors (37.8%) in the lower extremity, 62 (18%) in the pelvis, and 61 (17.7%) in the remaining trunk.

Table 1 Patient characteristics
Full size table

Performance of the DLM and the radiologists on the external test set

For comparison of diagnostic performance, the DLM as well as the four radiologists performed an analysis of the external test set. All results are summarized in Table 2.

Table 2 Performance parameters of the four readers and the deep learning algorithm for the differentiation between enchondromas vs. chondrosarcomas (G1-G3), enchondromas vs. acts as well as acts vs. High-Grade chondrosarcomas on the external dataset
Full size table

Differentiation enchondromas vs. chondrosarcomas (G1 - G3)

The DLM achieved a sensitivity, specificity, and accuracy of 97.0%, 32.4%, and 75.0% with an AUC of 0.88 for the differentiation between benign and malignant cartilaginous tumors on CT images. The overall sensitivity, specificity, accuracy of the residents’ readings were 92.4%, 47.1%, and 77.0%, and 92.4%, 67.6%, and 84.0%, respectively, with AUC values of 0.81 and 0.89. For the fellowship trained radiologist and the musculoskeletal attending radiologist, the overall sensitivity, specificity, accuracy were 98.5%, 64.7%, and 87.0%, and 97.0%, 76.5%, and 90.0%, respectively, with AUC values of 0.91 and 0.93. There was no significant difference found between the performance of the DLM and the musculoskeletal fellowship-trained radiologist (p = 0.19). Figure 2 shows the ROC analysis for the differentiation between benign and malignant chondroid tumors.

Fig. 2
figure 2

ROC-Analysis of the radiologists and the algorithm for the differentiation of enchondromas and chondrosarcomas (G1-G3). Reader 1 and 2 = residents; Reader 3 = fellowship trained radiologist; Reader 4 = musculoskeletal attending radiologist; ROC = receiver operating characteristic; AUC = area under the curve

Full size image

Differentiation enchondromas vs. ACTs

For the differentiation between enchondromas and ACTs the DLM achieved an overall sensitivity, specificity, and accuracy of 91.3%, 32.4%, and 56.1% with an AUC of 0.82. The overall sensitivity, specificity, and accuracy of the residents’ readings was 95.7%, 47.1%, and 66.7%, and 91.3%, 67.6%, and 77.2%, respectively, with AUC values of 0.81 and 0.87. The fellowship trained radiologist and the musculoskeletal attending radiologist achieved an overall sensitivity, specificity, and accuracy of 95.7%, 64.7%, and 77.2%, and 95.7%, 76.5%, and 84.2%, respectively, with AUC values of 0.87 and 0.92. There was no significant difference between the performance of the DLM and the four radiologic readers (p = 0.96, 0.33, 0.41, and 0.82). Figure 3 shows the ROC analysis for the differentiation between enchondromas and ACTs.

Fig. 3
figure 3

ROC-Analysis of the radiologists and the algorithm for the differentiation of enchondromas and atypical chondroid tumors. Reader 1 and 2 = residents; Reader 3 = fellowship trained radiologist; Reader 4 = musculoskeletal attending radiologist; ROC = receiver operating characteristic; AUC = area under the curve

Full size image

Differentiation ACTs vs. high-grade chondrosarcomas

Regarding the differentiation between ACTs and high-grade chondrosarcomas the DLM achieved a sensitivity, specificity, and accuracy of 53.5%, 60.9%, and 56.1% with an AUC of 0.64. The overall sensitivity, specificity, and accuracy of the residents’ readings were 55.8%, 52.2%, and 54.5%, and 55.8%, 52.2%, and 54.4%, respectively, with an AUC of 0.53 and 0.54. The overall sensitivity, specificity, and accuracy of the fellowship-trained radiologist and the musculoskeletal radiology attending were 95.3%, 30.4%, and 72.7%, and 86.0%, 26.1%, and 65.2%, respectively, with AUCs of 0.63 and 0.56. No significant difference was found between the performance of the DLM and the musculoskeletal attending radiologist (p = 0.26). Figure 4 shows the AUC curve for the differentiation between ACTs and enchondromas.

Fig. 4
figure 4

ROC-Analysis of the radiologists and the algorithm for the differentiation of atypical chondroid tumors and high-grade chondrosarcomas. Reader 1 and 2 = residents; Reader 3 = fellowship trained radiologist; Reader 4 = musculoskeletal attending radiologist; ROC = receiver operating characteristic; AUC = area under the ROC curve

Full size image

Inter-observer and intra-observer reliability

The inter-observer reliability was substantial to almost perfect for all criteria (κ = 0.76–1.00) and the intra-observer reliability was excellent (κ = 0.87–0.96), respectively. The dice score for the intrareader reliability was 0.97.

Discussion

The differentiation of chondroid tumors, especially between enchondromas and ACTs, is challenging on CT images yet crucial for determining appropriate therapeutic treatment and ensuring optimal patient outcomes. ACTs require a curettage as well as clinical and imaging follow-ups, whereas enchondromas typically do not require surgical intervention, unless symptomatic, and less frequent follow-ups. Therefore, this study aimed to develop a DLM for differentiating benign, intermediate and malignant chondroid tumors on CT images, comparing its diagnostic performance with that of radiologists. The DLM demonstrated a comparable performance to the experienced musculoskeletal radiologists.

To our knowledge, this is the first study showing the feasibility of using a DLM for the differentiation of chondroid tumors. In the past several studies have evaluated the diagnostic value of different imaging features for the differentiation of chondroid tumors [9,10,11]. Murphey et al. showed that cortical destruction, periosteal reaction, and endosteal scalloping (> 2/3 of cortical thickness) significantly increase the likelihood of a chondrosarcoma compared to an enchondroma [12]. Additionally, Crim et al. and Douis et al. identified tumor size, cortical breakthrough, periosteal reaction and matrix calcifications as key differentiating criteria [9, 10]. Nevertheless, differentiating enchondromas from ACTs radiologically remains challenging and varies with the radiologist’s expertise.

In our study, we used CT images for development of the DLM, although MRI can provide additional information. A recent study demonstrated that both CT and MRI show suggestive signs which can help to adequately differentiate enchondromas from ACTs in long bones. However, these features are rare, and a combination of CT and MRI features does not substantially improve the diagnostic performance [20].

Previous studies have shown the feasibility of using DLMs to differentiate between benign and malignant musculoskeletal tumors across various imaging modalities. Von Schacky et al., Liu at al., and He et al. developed CNN-based DLMs for the classification of benign, malignant, and intermediate primary bone tumors on radiographs in cohorts of 934, 643, and 1536 patients, respectively, achieving an AUC of up to 0.916 [15, 21, 22]. Several deep learning methods have been developed for detecting and classifying bone tumors using CT and MR images. For instance, a DLM based on radiomics in CT images and clinical parameters for discriminating between benign and malignant sacral tumors in 459 patients achieved an AUC of 0.83 [23]. Another DLM by Eweje et al. based on routine MR images and patient demographics achieved for the classification of benign and malignant bone lesions an AUC of 0.79 [24]. The performance of our DLM, with an AUC of 0.88 for differentiating between enchondromas and chondrosarcomas and an AUC of 0.82 for differentiating between enchondromas and ACTs, is comparable to those previous studies [23, 24].

In this study, a 2D CNN was employed, optimized via a mean squared error loss function, to facilitate the differential diagnosis of chondroid tumors utilizing CT data. The performance of the model, as demonstrated by the AUC scores, closely paralleled the diagnostic accuracy of experienced musculoskeletal radiologists. Such comparable efficacy across distinct patient cohorts underscores the potential of the model for generalization, an essential criterion for clinical translatability. The incorporation of this DLM into the diagnostic protocol may offer substantial clinical value, particularly as a supportive tool for radiologists with varying levels of expertise or experts in other fields aside from oncological or bone imaging. It can act as a supplementary computational consultation, thereby streamlining the diagnostic process and potentially impacting patient management strategies. Future research could explore model refinement through the integration of multimodal imaging variables and relevant clinical parameters.

While this study utilized a 2D CNN for tumor classification, future research could explore the potential advantages of 3D CNN architectures. A 3D CNN could leverage volumetric information from CT scans, potentially enhancing feature extraction and improving classification accuracy [25]. Prior studies have demonstrated that 3D deep learning models can be beneficial in various medical imaging tasks, particularly for segmenting and analyzing complex anatomical structures [26]. However, the increased computational cost, need for larger annotated datasets, and potential challenges in clinical interpretability must be considered. Furthermore, current radiological workflows predominantly rely on 2D slice-based assessments, making our approach more aligned with real-world clinical practice. Future studies could evaluate the feasibility of 3D CNNs in this setting and compare their performance against 2D approaches.

In addition to deep learning-based classification, advanced imaging modalities such as SPECT/CT radiomics have also shown potential in differentiating enchondromas from ACTs [27]. By extracting quantitative imaging features from functional imaging data, these approaches may complement conventional CT-based models. Future studies could explore the integration of SPECT/CT-based radiomics with deep learning models to enhance diagnostic accuracy.

One key finding of this study is the notable difference in classification performance between the differentiation of enchondromas from ACTs and high-grade chondrosarcomas, and the differentiation of ACTs from high-grade chondrosarcomas. This reflects a well-known challenge in musculoskeletal radiology, as these tumor types share overlapping imaging features on CT, particularly in terms of matrix mineralization and cortical involvement. Additionally, tumor heterogeneity within ACTs and high-grade chondrosarcomas may contribute to variability in their imaging presentation, further complicating classification. Previous studies have suggested that MRI, particularly contrast-enhanced and diffusion-weighted sequences, may offer additional diagnostic value in differentiating these entities. Future studies could explore multimodal deep learning approaches that integrate MRI data or employ radiomics-based feature extraction to enhance classification performance.

This study has limitations. Firstly, there has been a significant difference in the composition of the internal and external data set with a relatively low number of enchondromas in the external data set. Nevertheless, the overall number of patients in this study, to the authors knowledge, is the highest number of patients published in a single study [9, 10]. Secondly, in this study only patients with biopsy proven lesions were included, which may cause a bias towards more malignant looking lesions on the CT images. And lastly, our study did not include clinical parameters due to the retrospective design and large time frame of this study which was necessary to achieve such a big data set. Future studies may include clinical data as well as a combination of CT and MR images. And lastly, one of the main challenges in this study was the differentiation between ACTs and high-grade chondrosarcomas, where the model demonstrated lower performance compared to other classifications. This is consistent with prior studies reporting that even experienced musculoskeletal radiologists struggle with this distinction based on CT images alone. The overlapping imaging characteristics of ACTs and high-grade chondrosarcomas contribute to this difficulty. Future studies could explore the integration of MRI features or clinical parameters to enhance diagnostic accuracy.

Conclusions

In conclusion, the DLM reliably differentiates benign, intermediate, and malignant cartilaginous tumors, particularly enchondromas and ACTs, which are often difficult to distinguish based on CT imaging alone. However, its performance in differentiating ACTs from high-grade chondrosarcomas was more limited, in line with known diagnostic challenges.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Abbreviations

ACT:

Atypical cartilaginous tumors

AUC:

Area under the curve

CNN:

Convolutional neural network

DLM:

Deep learning model

PACS:

Picture archiving and communication systems

ROC:

Receiver operating characteristic

References

  1. Thorkildsen J, Taksdal I, Bjerkehagen B, Haugland HK, Børge Johannesen T, Viset T, Norum OJ, Bruland Ø, Zaikova O. Chondrosarcoma in Norway 1990–2013; an epidemiological and prognostic observational study of a complete national cohort. Acta Oncol (Stockholm Sweden). 2019;58(3):273–82.

    Article  Google Scholar 

  2. van Praag Veroniek VM, Rueten-Budde AJ, Ho V, Dijkstra PDS, Fiocco M, van de Sande MAJ. Incidence, outcomes and prognostic factors during 25 years of treatment of chondrosarcomas. Surg Oncol. 2018;27(3):402–8.

    Article  PubMed  Google Scholar 

  3. WHO Classification of Tumours Editorial Board eds. World Health Organization classification of soft tissue and bone tumours. 5th ed. Lyon. IARC. 2020.

    Google Scholar 

  4. Dierselhuis EF, Gerbers JG, Ploegmakers JJ, Stevens M, Suurmeijer AJ, Jutte PC. Local Treatment with Adjuvant Therapy for Central Atypical Cartilaginous Tumors in the Long Bones: Analysis of Outcome and Complications in One Hundred and Eight Patients with a Minimum Follow-up of Two Years. J Bone Joint Surg Am. 2016;98(4):303–13.

    Article  PubMed  Google Scholar 

  5. Nascimento D, Suchard G, Hatem M, de Abreu A. The role of magnetic resonance imaging in the evaluation of bone tumours and tumour-like lesions. Insights into imaging. 2014;5(4):419–40.

    Article  PubMed  PubMed Central  Google Scholar 

  6. Lalam R, Bloem JL, Noebauer-Huhmann IM, Wörtler K, Tagliafico A, Vanhoenacker F, Nikodinovska VV, Sanal HT, Woude HV, Papakonstantinou O, et al. ESSR Consensus Document for Detection, Characterization, and Referral Pathway for Tumors and Tumorlike Lesions of Bone. Semin Musculoskelet Radiol. 2017;21(5):630–47.

    Article  PubMed  Google Scholar 

  7. Laitinen MK, Stevenson JD, Parry MC, Sumathi V, Grimer RJ, Jeys LM. The role of grade in local recurrence and the disease-specific survival in chondrosarcomas. bone joint J. 2018;100–b(5):662–6.

    Article  PubMed  Google Scholar 

  8. Deckers C, Schreuder BH, Hannink G, de Rooy JW, van der Geest IC. Radiologic follow-up of untreated enchondroma and atypical cartilaginous tumors in the long bones. J Surg Oncol. 2016;114(8):987–91.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Crim J, Schmidt R, Layfield L, Hanrahan C, Manaster BJ. Can imaging criteria distinguish enchondroma from grade 1 chondrosarcoma? Eur J Radiol. 2015;84(11):2222–30.

    Article  PubMed  Google Scholar 

  10. Douis H, Parry M, Vaiyapuri S, Davies AM. What are the differentiating clinical and MRI-features of enchondromas from low-grade chondrosarcomas? Eur Radiol. 2018;28(1):398–409.

    Article  PubMed  Google Scholar 

  11. Flemming DJ, Murphey MD. Enchondroma and chondrosarcoma. Semin Musculoskelet Radiol. 2000;4(1):59–71.

    Article  CAS  PubMed  Google Scholar 

  12. Murphey MD, Flemming DJ, Boyea SR, Bojescul JA, Sweet DE, Temple HT. Enchondroma versus chondrosarcoma in the appendicular skeleton: differentiating features. Radiographics. 1998;18(5):1213–37. quiz 1244– 1215.

    Article  CAS  PubMed  Google Scholar 

  13. von Schacky CE, Sohn JH, Liu F, Ozhinsky E, Jungmann PM, Nardo L, Posadzy M, Foreman SC, Nevitt MC, Link TM, et al. Development and Validation of a Multitask Deep Learning Model for Severity Grading of Hip Osteoarthritis Features on Radiographs. Radiology. 2020;295(1):136–45.

    Article  Google Scholar 

  14. Ding Y, Sohn JH, Kawczynski MG, Trivedi H, Harnish R, Jenkins NW, Lituiev D, Copeland TP, Aboian MS, Mari Aparici C, et al. A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using (18)F-FDG PET of the Brain. Radiology. 2019;290(2):456–64.

    Article  PubMed  Google Scholar 

  15. von Schacky CE, Wilhelm NJ, Schäfer VS, Leonhardt Y, Gassert FG, Foreman SC, Gassert FT, Jung M, Jungmann PM, Russe MF, et al. Multitask Deep Learning for Segmentation and Classification of Primary Bone Tumors on Radiographs. Radiology. 2021;301(2):398–406.

    Article  Google Scholar 

  16. Song L, Li C, Tan L, Wang M, Chen X, Ye Q, Li S, Zhang R, Zeng Q, Xie Z, et al. A deep learning model to enhance the classification of primary bone tumors based on incomplete multimodal images in X-ray, CT, and MRI. Cancer Imaging. 2024;24(1):135.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Yildirim M, Yildirim H. CT radiomics-based machine learning model for differentiating between enchondroma and low-grade chondrosarcoma. Med (Baltim). 2024;103(33):e39311.

    Article  Google Scholar 

  18. Hong R, Li Q, Ma J, Lu C, Zhong Z. Computed tomography-based radiomics machine learning models for differentiating enchondroma and atypical cartilaginous tumor in long bones. Rofo. 2024.

  19. Powers DM. Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation. J Mach Learn Technol. 2011;2(1);37–63.

  20. Gassert FG, Breden S, Neumann J, Gassert FT, Bollwein C, Knebel C, Lenze U, von Eisenhart-Rothe R, Mogler C, Makowski MR et al. Differentiating Enchondromas and Atypical Cartilaginous Tumors in Long Bones with Computed Tomography and Magnetic Resonance Imaging. Diagnostics (Basel) 2022;12(9).

  21. Liu R, Pan D, Xu Y, Zeng H, He Z, Lin J, Zeng W, Wu Z, Luo Z, Qin G, et al. A deep learning-machine learning fusion approach for the classification of benign, malignant, and intermediate bone tumors. Eur Radiol. 2022;32(2):1371–83.

    Article  PubMed  Google Scholar 

  22. He Y, Pan I, Bao B, Halsey K, Chang M, Liu H, Peng S, Sebro RA, Guan J, Yi T, et al. Deep learning-based classification of primary bone tumors on radiographs: A preliminary study. EBioMedicine. 2020;62:103121.

    Article  PubMed  PubMed Central  Google Scholar 

  23. Yin P, Mao N, Chen H, Sun C, Wang S, Liu X, Hong N. Machine and Deep Learning Based Radiomics Models for Preoperative Prediction of Benign and Malignant Sacral Tumors. Front Oncol. 2020;10:564725.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Eweje FR, Bao B, Wu J, Dalal D, Liao WH, He Y, Luo Y, Lu S, Zhang P, Peng X, et al. Deep Learning for Classification of Bone Lesions on Routine MRI. EBioMedicine. 2021;68:103402.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Ilesanmi AE, Ilesanmi TO, Ajayi BO. Reviewing 3D convolutional neural network approaches for medical image segmentation. Heliyon. 2024;10(6):e27398.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Soleimani P, Farezi N. Utilizing deep learning via the 3D U-net neural network for the delineation of brain stroke lesions in MRI image. Sci Rep. 2023;13(1):19808.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  27. Yoon H, Choi WH, Joo MW, Ha S, Chung YA. SPECT/CT Radiomics for Differentiating between Enchondroma and Grade I Chondrosarcoma. Tomography. 2023;9(5):1868–75.

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

Not applicable.

Funding

Open Access funding enabled and organized by Projekt DEAL.

This work was funded by grants from the German Society of Musculoskeletal Radiology (Deutsche Gesellschaft für muskuloskelettale Radiologie; DGMSR) and by the Seed Grant from the International Skeletal Society.

Author information

Author notes

  1. Felix G. Gassert, Daniel Lang, Nina Hesse, Florian T. Gassert and Alexandra S. Gersing contributed equally to this work.

Authors and Affiliations

  1. Department of Radiology and Biomedical Imaging, University of California San Francisco, 505 Parnassus Ave, San Francisco, CA, 94143, USA

    Felix G. Gassert & Alexandra S. Gersing

  2. Department of Radiology, Klinikum Rechts der Isar, School of Medicine and Health, Technical University of Munich, Ismaninger Strasse 22, 81675, Munich, Germany

    Felix G. Gassert, Luca Kohll & Florian T. Gassert

  3. Institute of Radiation Medicine, Helmholtz Zentrum München, Ingolstädter Landstraße 1, 85764, Neuherberg, Germany

    Daniel Lang, Jan C. Peeken & Stefan Bartzsch

  4. Department of Radiation Oncology, School of Medicine and Health, Klinikum Rechts der Isar, Technical University of Munich, Ismaninger Strasse 22, 81675, Munich, Germany

    Daniel Lang, Jan C. Peeken & Stefan Bartzsch

  5. Orthopaedic Oncology, Department of Orthopaedics and Trauma Surgery, Musculoskeletal University Center Munich (MUM), LMU Klinikum, University Hospital, LMU Munich, Marchioninistraße 15, 81377, Munich, Germany

    Hans Roland Dürr & Alexander Klein

  6. Department of Orthopaedics and Sports Orthopaedics, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Ismaninger Strasse 22, 81675, Munich, Germany

    Florian Hinterwimmer & Carolin Knebel

  7. Department of Radiology, LMU University Hospital, LMU Munich, Marchioninistraße 13, 80337, Munich, Germany

    Nina Hesse, Johanna Luitjens & Stefan Weissinger

  8. Department of Pathology, Klinikum Rechts der Isar, School of Medicine and Health, Technical University of Munich, Ismaninger Strasse 22, 81675, Munich, Germany

    Carolin Mogler

  9. Department of Neuroradiology, LMU University Hospital, LMU Munich, Marchioninistraße 13, 80337, Munich, Germany

    Alexandra S. Gersing

  10. Institute for AI and Informatics in Medicine, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Munich, Germany

    Florian Hinterwimmer

Authors
  1. Felix G. Gassert
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Daniel Lang
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Nina Hesse
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. Hans Roland Dürr
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Alexander Klein
    View author publications

    You can also search for this author inPubMed Google Scholar

  6. Luca Kohll
    View author publications

    You can also search for this author inPubMed Google Scholar

  7. Florian Hinterwimmer
    View author publications

    You can also search for this author inPubMed Google Scholar

  8. Johanna Luitjens
    View author publications

    You can also search for this author inPubMed Google Scholar

  9. Stefan Weissinger
    View author publications

    You can also search for this author inPubMed Google Scholar

  10. Jan C. Peeken
    View author publications

    You can also search for this author inPubMed Google Scholar

  11. Carolin Mogler
    View author publications

    You can also search for this author inPubMed Google Scholar

  12. Carolin Knebel
    View author publications

    You can also search for this author inPubMed Google Scholar

  13. Stefan Bartzsch
    View author publications

    You can also search for this author inPubMed Google Scholar

  14. Florian T. Gassert
    View author publications

    You can also search for this author inPubMed Google Scholar

  15. Alexandra S. Gersing
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

Conceptualization, F.G.G, J.L, and A.S.G.; methodology, F.T.G.; software, D.L.; validation, F.G.G., J.L, F.T.G. and A.S.G.; formal analysis, D.L.; investigation, L.K.; resources, F.H., H.R.D., C.M., and C.K.; data curation, L.K., S.W.; writing—original draft preparation, F.G.G:, D. L, N.H.; writing—review and editing, F.G.G, D.L., N.H., H.R.D., A.K., L.K., F.H., J.L., S.W., J.C.P., C.M. C.K., S.B., F.T.G., A.S.G.; visualization, J.L.; supervision, F.T.G., A.S.G.; project administration, F.T.G., A.S.G; funding acquisition, A.S.G. All authors have read and agreed to the published version of the manuscript.

Corresponding authors

Correspondence to Florian T. Gassert or Alexandra S. Gersing.

Ethics declarations

Ethics approval and consent to participate

Approval of the Institutional Review Boards had been obtained prior to this study (Institutional Review Board of the Technical University of Munich, approval number 393/20, and the Ludwig-Maximilian University of Munich, approval number 21–0282). Written informed consent was waived for this retrospective analysis of routinely acquired imaging and clinical data by the Institutional Review Board of the Technical University of Munich and the Institutional Review Board of the Ludwig-Maximilian University of Munich. All analyses are in line with the declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Gassert, F.G., Lang, D., Hesse, N. et al. A deep learning model for classification of chondroid tumors on CT images. BMC Cancer 25, 561 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12885-025-13951-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12885-025-13951-1

Keywords

BMC Cancer

ISSN: 1471-2407

Contact us