Fusion of Deep Transfer Learning and Radiomics in MRI-Based Prediction of Post-Surgical Recurrence in Soft Tissue Sarcoma

Abstract

Soft  tissue sarcomas (STS) are heterogeneous malignancies with high recurrence rates (33–39%) post-surgery, necessitating improved prognostic tools. This study proposes a fusion model integrating deep transfer learning and radiomics from MRI to predict postoperative STS recurrence. Axial T2-weighted fat-suppressed imaging (T₂WI) of 803 STS patients from two institutions was retrospectively collected and divided into training (n = 527), internal validation (n = 132), and external validation (n = 144) cohorts. Tumor segmentation was performed using the SegResNet model within the Auto3DSeg framework. Radiomic features and deep learning features were extracted. Feature selection employed LASSO regression, and the deep learning radiomic (DLR) model combined radiomic and deep learning signatures. Using the features, nine models were constructed based on three classifiers. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, negative predictive value, and positive predictive value were calculated for performance evaluation. The SegResNet model achieved Dice coefficients of 0.728 after refinement. Recurrence rates were 22.8% (120/527) in the training, 25.0% (33/132) in the internal validation, and 32.6% (47/144) in the external validation cohorts. The DLR model (ExtraTrees) demonstrated superior performance, achieving an AUC of 0.818 in internal validation and 0.809 in external validation, better than the radiomic model (0.710, 0.612) and the deep learning model (0.751, 0.667). Sensitivity and specificity ranged from 0.702 to 0.976 and 0.732 to 0.830, respectively. Decision curve analysis confirmed superior clinical utility. The DLR model provides a robust, non-invasive tool for preoperative STS recurrence prediction, enabling personalized treatment decisions and postoperative management.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-025-01653-w.

Introduction

Soft tissue sarcomas (STS) are a heterogeneous group of mesenchymal malignancies with diverse histopathological subtypes and variable clinical outcomes, accounting for 1% of the overall incidence of adult cancer [1]. Despite advances in surgical resection and multimodal therapy, a significant proportion of STS patients, ranging from 33 to 39%, experience local recurrence [2]. Traditional risk assessment relies on clinical and pathological parameters, such as tumor size, histological grade, and resection margin status, yet these factors alone have limited predictive power [3]. Once recurrence occurs, the prognosis of patients deteriorates significantly, severely compromising their physical and mental well-being, as well as their overall survival. Consequently, there is an urgent need for a more refined predictive model. This model would empower clinicians to implement more aggressive therapeutic strategies, thereby reducing the likelihood of recurrence [4].

Radiomics, which involves extracting high-dimensional quantitative features from medical images, has shown promise in oncologic imaging by enabling non-invasive tumor characterization and prognosis prediction [5]. Several studies have demonstrated that computed tomography and magnetic resonance imaging (MRI)-based radiomics can predict STS recurrence and survival outcomes [6–8]. However, the reproducibility and generalizability of radiomic models remain challenges due to variations in imaging protocols, segmentation strategies, and feature extraction methods [9].

Deep learning (DL), particularly convolutional neural networks, has emerged as a powerful tool for automated feature extraction and pattern recognition in medical imaging [10]. Previous studies have validated deep learning radiomic nomograms for STS recurrence prediction, demonstrating superior prognostic capabilities compared to traditional clinical models [11]. By integrating with radiomics, DL-based approaches capture complex, high-order imaging patterns beyond handcrafted features, enhancing model performance [12]. Deep transfer learning, a subset of transfer learning that leverages pre-trained models on large datasets, is particularly valuable for improving predictive accuracy in small, heterogeneous datasets such as STS [13].

However, existing studies often suffer from single-center designs, small sample sizes, and inconsistent external validation, limiting clinical translation [14]. Furthermore, while contrast-enhanced MRI enhances tumor characterization, its invasiveness and variability across institutions complicate widespread application [15].

In this study, we developed a multicenter DL-radiomics framework using big data from non-contrast MRI to predict postoperative STS recurrence. By integrating radiomic features and DL-derived patterns to overcome the limitations of traditional methods, leveraging a large-scale dataset helps overcome issues of model applicability and the heterogeneity across different centers. Offering a non-invasive and reproducible model, our approach promises reliable, personalized risk stratification. This provides crucial support for adjuvant treatment decision-making, thereby effectively prolonging patient survival.

Method

Patients

This study was approved by the Ethics Committee of our hospital and complied with the Declaration of Helsinki. Informed consent was waived because the present study is retrospective.

In this study, 659 patients were pathologically confirmed as STS in our hospital from April 2012 to November 2022, and 144 patients were confirmed in another center from January 2007 to May 2018. They were recruited for eligibility according to the inclusion and exclusion criteria (Fig. 1). Pathological, clinical, and imaging data were collected. The medical records and baseline tumor characteristics were reviewed by a trained radiologist with 3 years of experience. Demographic information was obtained from patients'electronic medical records, including age at the time of imaging, gender, and pathology type.

Fig. 1.

Flowchart of included patients in our study

This study divided the samples into three cohorts: a training cohort, comprising 80% of the data, an internal validation cohort, consisting of the remaining 20%, and an external validation cohort.

Each patient underwent at least one MRI follow-up within three years after surgery. The presence of a mass detected by MRI was considered indicative of recurrence, while no abnormalities detected over three years were deemed non-recurrent.

Image Acquisition

The MRI examinations were performed using 3.0 T Discovery 750 GE Healthcare, 3.0 T Discovery 750 W GE Healthcare, Philips Ingenia 3.0 T, 1.5 T uMR 588 United Imaging, and 8-channel phased array torso coils in both institutions. All patients were positioned supine and feet-first. Images used in this study included axial T2-weighted images (T₂WI) with fast recovery fast spin-echo with fat saturation. Raw digital imaging and communications in medicine (DICOM) files were de-identified and transferred from each institution’s picture archiving and communication system to a centralized, encrypted server for analysis, compliant with institutional data privacy policies.

Image Pre-Processing and Automatic Tumor Segmentation

All DICOM images were extracted and converted to the NiFTI format prior to the analysis using the dcm2niix software [16]. Subsequently, all images were resampled to 1 × 1 × 10 mm³. The images were normalized using min–max normalization to a range of 0–1, since MRI acquisition settings can result in substantial variations in signal intensities of the STS. The ground truth labels were established through manual segmentation for the dataset. In our study, a more refined approach was adopted for the segmentation process. Initially, the volume of interest (VOI) of 200 samples in the training cohort was manually delineated using 3D Slicer. Tumor boundaries were traced slice-by-slice along the margins. Where tumor boundaries were ambiguous or demarcation from edema proved challenging, supplementary sequences (e.g., contrast-enhanced or diffusion-weighted imaging) were utilized for clarification. Indeterminate cases were resolved through discussion with senior attending radiologists. This task was carried out by two experienced radiologists, and any discrepancies in their annotations were resolved by a senior radiologist with over 20 years of experience. These meticulously hand-drawn VOIs were then utilized to train the Auto3DSeg framework (https://github.com/Project-MONAI/tutorials/tree/main/auto3dseg) derived from MONAI, which contains the SegResNet models [17]. The Auto3DSeg algorithm first analyzed the intensity, size, spacing, and other characteristics of these 200 training images. Based on the analysis, it automatically determined hyperparameters such as training epochs, input image size, and batch size for the segmentation model. The performance of the segmentation model was evaluated using the Dice coefficient [18], and then it performed automatic segmentation of the tumor regions in the remaining 603 patients. Experienced radiologists reviewed the segmented images, fine-tuned the biased parts, and re-input the corrected images into the segmentation model for training, significantly enhancing the efficiency and accuracy of tumor region delineation. One month later, 30 patients in the training sequence were randomly selected and re-delineated by two experienced radiologists. The reproducibility of the radiomic features was evaluated using the intraclass correlation coefficient (ICC).

Radiomic Features Extraction

A total of 1015 radiomic features were extracted based on PyRadiomics (version 1.3.0; Computational Imaging and Bioinformatics Lab, Harvard Medical School), which is an open-source software platform dedicated to feature extraction for medical images [19]. The VOI with manually delineated tumor contours was imported into this platform, and then quantitative features were extracted through a series of radiomic algorithms. Details of the radiomic features can be found in Supplementary Fig. S1.

Deep Learning Features Extraction

In our processing pipeline, for each patient, the slice displaying the largest volume of interest was chosen as the representative image. To streamline the analysis and minimize background noise, the smallest enclosing cuboid was extracted. The intensity distribution across the images was standardized using z-score normalization. These normalized images were then used as inputs for our DL model. During the training phase, real-time data augmentation strategies such as random cropping and flipping were implemented to enhance model robustness. For the test images, preprocessing was confined to normalization only, ensuring consistency in the evaluation process.

In this study, we explored the performance of prominent networks such as DenseNet121 to enhance the performance of traditional CNN-based models. Additionally, we conducted comparative analyses of these models to identify the most suitable algorithm for our specific research requirements. In our study, to ensure the model’s effectiveness across various patient populations with notable variability, we implemented transfer learning. This process involved initializing the model with pre-trained weights from the ImageNet database, enhancing its adaptability to diverse datasets. A critical aspect of our approach was the meticulous adjustment of the learning rate to foster better generalization across datasets. To improve the transparency and interpretability of the model’s decision-making process, we employed gradient-weighted class activation mapping (Grad-CAM) to visualize the models. The gradient information from the last convolutional layer of the DenseNet121 was used for weighted fusion to generate a class activation map that emphasizes the critical regions in the target classification image [20].

Feature Selection and Development of Predictive Models

First, the radiomic features were normalized using the z-score method. Additionally, for features exhibiting high repeatability, Spearman’s rank-correlation coefficient was used to calculate the correlation between features. If the correlation coefficient between two features exceeded 0.9, only one of the features was retained. Subsequently, the least absolute shrinkage and selection operator (LASSO) logistic regression (LR) was employed to remove irrelevant and redundant features [21]. The optimal λ value was determined through tenfold cross-validation, where the value providing the minimum cross-validation error was selected, details of the coefficients and MSE of tenfold cross-validation in Supplementary Fig. S2. For constructing the deep learning radiomic (DLR) model, we employed a pre-fusion algorithm that combines the DL features with radiomic features. Then, the selected radiomic features, DL features, and DLR features were used as input, respectively, for the k-nearest neighbors (k-NN), ExtraTrees, and multi-layer perceptron (MLP) classifiers to construct the radiomic (rad) model, the DL model, and the DLR model [22]. Univariable logistic regression analysis was performed in the training cohort to screen clinical factors associated with STS recurrence. Significant variables were subsequently incorporated into multivariable logistic regression analysis. Using a backward stepwise selection approach with p < 0.05 as the retention criterion, independent clinical predictors of STS recurrence were ultimately identified. A multivariable clinical prediction model was thus constructed based on these factors. To address dataset imbalance and improve the model’s sensitivity to minority-class predictions, we applied the synthetic minority over-sampling technique (SMOTE) during model training. The hyperparameters of classifiers are shown in Table S1. The study workflow is illustrated in Fig. 2.

Fig. 2.

The workflow of this study. Tumors were contoured manually on T₂WI, and the VOI enclosing the tumor was cropped as the predictive model input. The DL features and radiomic features were separately extracted from eight phases of MR images. The LASSO logistic regression algorithm was applied to select two types of features and construct a rad model, a DL model, and a DLR model. Results analysis was performed through ROC curves, DCA, and the DeLong test

Metrics

The diagnostic efficacy of our three models was rigorously evaluated in the test cohort through the construction of ROC curves to assess their discriminative ability. Calibration of the model was analyzed using calibration curves. Additionally, decision curve analysis (DCA) was employed to appraise the clinical utility of our predictive models, facilitating an understanding of the potential benefits in a clinical context.

Statistical Analysis

The normality of clinical features was verified using the Shapiro–Wilk test. Continuous variables were subjected to the t-test or the Mann–Whitney U test based on their distribution characteristics. Categorical variables were examined using chi-square (χ²) tests. The area under the receiver operating characteristic (ROC) curve (AUC) of the predictive models was compared, with sensitivity, specificity, positive predictive value, and negative predictive value calculated. The statistical analyses were performed using R software, version 3.5.1 (https://www.rproject.org/). All statistical tests were two-sided, and p < 0.05 indicated significance.

Results

Clinical Characteristics and Automatic Tumor Segmentation

We screened 659 pathologically confirmed STS patients in our hospital and 144 pathologically confirmed STS patients in another center. The patients were randomly allocated to three cohorts: training (n = 527), internal validation (n = 132), and external validation (n = 144), with the mean age of 44.3 ± 19.8, 45.6 ± 19.8, and 57.4 ± 16.3 years, respectively. Among the 803 patients, 200 (24.9%) had recurrence, and 603 (75.1%) did not have recurrence. The mean age was 45.6 ± 20.3 years in the recurrence group and 50.6 ± 18.0 years in the no-recurrence group. Patients in the non-recurrence group were significantly older than those in the recurrence group (p = 0.001). The mean age in the external validation cohort was higher than in the internal validation cohort and the training cohort, with a statistically significant difference (P < 0.001). In addition, the recurrence rate in the external validation cohort was statistically higher than that of the training and internal validation cohorts (P = 0.022). The main pathological subtypes of soft tissue sarcoma include: malignant peripheral nerve sheath tumor, leiomyosarcoma, angiosarcoma, extraskeletal sarcoma, undifferentiated sarcoma, liposarcoma, rhabdomyosarcoma, synovial sarcoma, fibrosarcoma, and alveolar soft part sarcoma. Angiosarcoma, extraskeletal sarcoma, alveolar soft part sarcoma, and leiomyosarcoma are less common types and will be grouped together as “others” for statistical purposes. The characteristics of the patients in the cohorts are summarized in Table 1 and Table S2.

Table 1.

Baseline characteristics of patients in cohorts

Characteristics	Recurrence	No recurrence	P
Age/years	45.6 ± 20.3	50.6 ± 18.0	0.001
Sex/n(%)			0.032
Female	302 (50.1)	82 (41.0)
Male	301 (49.9)	118 (59.0)
Type/n(%)			0.373
Malignant peripheral nerve sheath tumor	46 (7.6)	24 (12.0)
Rhabdomyosarcoma	50 (8.3)	16 (8.0)
Synovial sarcoma	81 (13.4)	19 (9.5)
Undifferentiated sarcoma	203 (33.7)	67 (33.5)
Fibrosarcoma	61 (10.1)	22 (11.0)
Liposarcoma	112 (18.6)	40 (20.0)
Others	50 (8.3)	12 (6.0)

The automatic segmentation SegResNet models were trained for 1000 epochs and achieved a dice coefficient of 0.671 in the training cohort. The SegResNet was utilized to automatically segment the tumor region for the remaining 603 patients. Subsequently, experienced radiologists reviewed the segmented images, refined the biased regions, and reintegrated the corrected images into the segmentation model for further training. This approach greatly enhanced the efficiency of manual VOI delineation. Finally, we re-evaluated the performance of the SegResNet model in the training cohort with a dice coefficient of 0.728, which could be used in the future for better performance.

Feature Extraction and Selection for Model Construction

A total of 1015 radiomic features and 1024 DL features in the eight phases were originally extracted from the segmentations. The extracted radiomic features consisted of seven feature groups, with first-order statistical features (n = 198), shape-based features (n = 14), and texture features. To investigate the DL model’s recognition abilities on various samples, we used the Grad-CAM technique for visualization in Fig. 3.

Fig. 3.

Illustrates Grad-CAM’s application, highlighting the activations in the final convolutional layer relevant to cancer type prediction

Based on the LASSO logistic regression algorithm, 11 radiomic features and 6 DL features with nonzero correlation coefficients of STS recurrence were initially selected for the construction of the radiomic model and the DL model. The construction of the DLR model was based on combined features, which were ultimately screened from the above 10 radiomic and 7 DL features using the LASSO algorithm. Supplementary Fig. S3 illustrates the coefficient values in the final selected non-zero features for the rad model, the DL model, and the DLR model.

Performance and Comparisons of Three Predictive Models

In the training cohort, the DLR ExtraTrees model achieved the highest AUC of 0.969 (95% CI 0.956–0.981), with an accuracy of 0.861, a sensitivity of 0.976, and a specificity of 0.830 for predicting recurrence of STS. In the external validation cohort, the DLR ExtraTrees model achieved the highest AUC of 0.809 (95% CI 0.734–0.882) and a preserved high accuracy of 0.722, with balanced sensitivity and specificity of 0.702 and 0.732. In the internal validation cohort, the DLR ExtraTrees model achieved the highest AUC of 0.818 (95% CI 0.742–0.895) and a preserved high accuracy of 0.788, with balanced sensitivity and specificity of 0.727 and 0.808.

With the utilization of the DL features, the performance of our DLR model was significantly better than that of the rad and DL models. The DLR model exhibited satisfactory predictive power in both the training and two validation cohorts (AUC = 0.969, 0.818, and 0.808, respectively) and performed better than the rad and DL models in the two cohorts. A detailed comparison of the three models is presented in Table 2. The results of ROC analysis in the training and validation cohorts are displayed in Fig. 4. DeLong’s test for multiple ROC curves showed that the DLR model for the ExtraTrees classifier achieved superior discriminative performance compared to both the rad model (P < 0.001) and the DL model (P = 0.005). More results are presented in Fig. S4 and S5. Based on multivariable logistic regression analysis incorporating age and sex, the clinical model demonstrated significantly lower AUC values in both the internal and external validation cohorts (0.478 and 0.526, respectively) compared to the DLR model (Table S3).

Table 2.

The comparison of performance of the Rad model, the DL model, and the DLR model

Train	ExtraTrees			k-NN			MLP
	Rad	DL	DLR	Rad	DL	DLR	Rad	DL	DLR
Accuracy	0.812	0.816	0.861	0.888	0.896	0.899	0.636	0.657	0.645
AUC	0.881	0.921	0.969	0.921	0.932	0.923	0.729	0.657	0.781
95% CI	0.848, 0.914	0.898, 0.945	0.956, 0.981	0.895, 0.946	0.908, 0.956	0.897, 0.948	0.678, 0.779	0.604, 0.710	0.738, 0.824
Sensitivity	0.783	0.900	0.967	0.508	0.542	0.558	0.708	0.567	0.892
Specificity	0.821	0.791	0.830	1.000	1.000	1.000	0.614	0.683	0.572
PPV	0.563	0.560	0.627	1.000	1.000	1.000	0.351	0.345	0.381
NPV	0.928	0.964	0.988	0.873	0.881	0.885	0.877	0.842	0.947
Precision	0.563	0.560	0.627	1.000	1.000	1.000	0.351	0.345	0.381
I-Val
Accuracy	0.697	0.629	0.788	0.720	0.583	0.598	0.598	0.705	0.636
AUC	0.710	0.751	0.818	0.553	0.643	0.650	0.734	0.812	0.751
95% CI	0.612, 0.807	0.669, 0.833	0.741, 0.895	0.436, 0.670	0.542, 0.745	0.550, 0.750	0.640, 0.828	0.731, 0.893	0.658, 0.844
Sensitivity	0.667	0.909	0.727	0.061	0.545	0.606	0.848	0.848	0.879
Specificity	0.707	0.535	0.808	0.939	0.596	0.596	0.515	0.657	0.556
PPV	0.431	0.395	0.558	0.250	0.310	0.333	0.368	0.452	0.397
NPV	0.864	0.946	0.899	0.750	0.797	0.819	0.911	0.929	0.932
Precision	0.431	0.395	0.558	0.250	0.310	0.333	0.368	0.452	0.397
E-Val
Accuracy	0.632	0.611	0.722	0.674	0.667	0.618	0.674	0.535	0.590
AUC	0.612	0.667	0.809	0.514	0.639	0.536	0.649	0.600	0.624
95% CI	0.515, 0.708	0.576, 0.758	0.734, 0.883	0.412, 0.616	0.542, 0.735	0.438, 0.635	0.555, 0.742	0.504, 0.696	0.528, 0.719
Sensitivity	0.553	0.766	0.702	0.000	0.319	0.170	0.489	0.745	0.660
Specificity	0.670	0.536	0.732	1.000	0.835	0.835	0.763	0.433	0.557
PPV	0.448	0.444	0.559	0.000	0.484	0.333	0.500	0.389	0.419
NPV	0.756	0.825	0.835	0.674	0.717	0.675	0.755	0.778	0.771
Precision	0.448	0.444	0.559	0.000	0.484	0.333	0.500	0.389	0.419

I-val, internal validation cohort; E-val, validation cohort; AUC, area under the receiver operating characteristic curve; CI, confidence interval; ACC, accuracy; PPV, positive predictive value; NPV, negative predictive value; Rad, radiomic model; DL, deep learning model; DLR, deep learning radiomic model

Fig. 4.

The ROC curves of the rad model, DL model, and DLR model in the training cohort (a), internal validation cohort (b), and external validation cohort (c). The DLR model demonstrated significantly higher AUCs in the training and validation cohorts (AUC of 0.969, 0.818, and 0.809) than the rad model (AUC of 0.881, 0.710, and 0.610) and the DL model (AUC of 0.921, 0.751, and 0.667)

Additionally, the calibration curve and DCA of the DLR model are displayed in Fig. 5. The calibration curve of the DLR model demonstrates a strong agreement between predictions and observations in the validation cohort. The DCA of the DLR model demonstrates superior clinical benefits for predicting recurrence. Compared to scenarios where no prediction model is used, the DLR demonstrates a significant advantage in prediction probability. When compared to the rad and DL models, the DLR also shows its superiority.

Fig. 5.

Calibration curves of different signatures (a) and different signatures’ decision curve (b) in the external validation cohort

Discussion

This multicenter study extracted critical DL and radiomic features from preoperative MRI plain-scan T₂WI. By integrating DL and radiomic features, we developed and validated a combined model that achieved a higher AUC of 0.808 in an independent external validation cohort compared to two single-modality models using either feature type alone. This confirms the superior performance of the DLR model in predicting postoperative recurrence of STS. To the best of our knowledge, this study represents the largest multicenter dataset for STS to date, demonstrating enhanced generalizability and reproducibility while providing a robust and interpretable tool for personalized treatment decisions and postoperative management.

Our DLR model ExtraTrees classifier exhibited superior performance, achieving a training AUC of 0.969, which significantly outperformed both standalone radiomic models and pure DL architectures. The external validation results (AUC = 0.809) surpassed those reported by Liu [11] (AUC = 0.746), utilizing transfer learning to ensure robust generalization across different imaging protocols. The superior performance of the ExtraTrees classifier observed in this study could be attributed to its distinctive capacity to model complex non-linear relationships through recursive partitioning of the feature space, which appears particularly advantageous for recurrence prediction tasks. This machine learning approach demonstrates strong clinical interpretability, as its hierarchical rule-based structure aligns well with clinicians'diagnostic reasoning processes [23]. Casale et al. [8] demonstrated that radiomics and formal methods achieved a sensitivity of 0.81, specificity of 0.67, and accuracy of 0.74 in predicting recurrence and metastasis. In this study, the sensitivity (0.727) of the internal validation cohort was lower than that reported by Casale, while both the specificity (0.828) and accuracy (0.788) exceeded their results. Formal methods can enhance both reliability and security in computer systems, offering valuable support for personalized medicine. It is particularly noteworthy that our study benefits from a larger multicenter sample size and the application of a fusion model, significantly enhancing the reliability of the findings. Schmitz et al. [24] demonstrated that radiomics can predict high-grade soft tissue sarcoma with an area under the curve (AUC) of 0.97. Given the established strong association between high-grade tumors and recurrence, this finding aligns with the results of our study. The lower performance observed in our test set compared to Schmitz's study may be attributable to the absence of contrast-enhanced data and inherent heterogeneity within our multicenter dataset. However, our approach, utilizing DL and multicenter data, enhances clinical applicability while improving model generalizability.

The observed degradation in external validation metrics relative to internal validation likely originates from inter-institutional heterogeneity in data distributions—a well-documented phenomenon in multicenter medical imaging studies that underscores fundamental challenges in model generalizability across heterogeneous clinical environments [25, 26]. Additional concerns regarding model applicability arise from the absence of integrated clinical data repositories and inherent selection biases associated with retrospective study designs. Nevertheless, DL methodologies offer a promising framework to mitigate these limitations. For future clinical translation, our model architecture could serve as a domain-adaptive foundation, whereby center-specific data would undergo institution-targeted transfer learning to develop customized predictive models. This paradigm would deliver actionable preoperative intelligence to clinicians while maintaining flexibility for subsequent integration of multimodal clinical data, thereby enabling institution-specific customization across diverse healthcare settings.

Although significant differences exist in age distribution between the training and external validation cohorts, as well as in age and sex distribution between the recurrence and non-recurrence groups, this discrepancy did not influence the predictive performance of the model because it is not a true biological mechanism, and the findings from our clinical model also confirm this. The higher recurrence rate observed in the external validation cohort compared to the training and internal validation cohorts may be attributed to variations in treatment practices across institutions. This discrepancy could theoretically reduce the predictive accuracy of the model, but our results still demonstrated robust performance through the application of the smoothing technique. We only selected T₂WI for analysis because the large cohort size of patients was sufficient for training and testing the model, and T₂WI imaging demonstrated small differences across different scanners. Moreover, this approach better enhances clinical applicability in resource-constrained settings where advanced imaging protocols may be inaccessible. To justify the clinical usefulness, we also applied DCA curve analysis, showing that the DLR model for early recurrence prediction had more benefits than either the treat-all or the treat-none strategy across the full range of reasonable threshold probabilities. Additionally, we did not use a nomogram to demonstrate survival probability due to several limitations, particularly its lack of clinical significance and its inability to effectively support data-driven clinical decision-making [27].

Our key innovation in this study is that transfer learning via pre-trained DenseNet121 circumvented overfitting risks inherent to small medical imaging datasets, aligning with emerging evidence that DLR enhances prognostic modeling in rare tumors. The DLR model enables preoperative identification of high-risk STS patients, outperforming traditional clinicopathological criteria with an AUC not exceeding 0.74 [11, 28]. This could guide neoadjuvant therapy intensification or wider resection margins in high-risk cases, potentially reducing recurrence rates. Conversely, the high NPV (0.929) supports de-escalation of surveillance for low-risk patients, minimizing unnecessary interventions. Importantly, this MRI-based approach avoids invasive biopsies and enhances accessibility compared to molecular profiling, aligning with precision oncology goals [29]. Besides, the automated 3D segmentation framework (SegResNet) achieved Dice coefficients of 0.671–0.728, reducing manual delineation time.

Despite these advances, limitations warrant attention. First, while external validation incorporated multicenter data, the retrospective design inherently risks selection bias. Prospective validation across diverse, prospectively enrolled cohorts is essential to confirm clinical applicability and mitigate potential confounding factors. Second, although rigorous standardization and feature screening substantially reduced inter-scanner variability, residual effects of protocol differences may persist. This heterogeneity likely contributed to the observed performance drop during external validation and remains a challenge for broad model transferability. Widespread adoption of harmonized protocols is critical for future clinical deployment. Therefore, we note that while the large test set provides robust stability, future work conducted under sufficient computational resources would benefit from employing more intensive resampling techniques, such as k-fold CV, to yield refined performance intervals and potentially reduce residual variance associated with a single partition. Last, our model exclusively utilized imaging-derived features, omitting clinical and STS subtype-related data known to influence recurrence risk. Specific subtypes might behave differently from others. However, due to the rarity of most of the subtypes in STS, an analysis of the subtypes separately is difficult. Future work must integrate multimodal data to enhance predictive completeness and clinical utility.

In conclusion, the prediction model we constructed based on DL combined with radiomic features has been proven to be a powerful and externally validated tool for predicting the recurrence of STS. This offers an unprecedented opportunity to optimize the formulation of treatment protocols for patients and potentially improve patient survival. More research is required in the future to validate the application of this model in clinical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

STS

Soft tissue sarcomas

AUC

Area under the receiver operating curve

DCA

Decision curve analysis

MRI

Magnetic resonance imaging

DL

Deep learning

VOI

Volume of interest

ICC

Intraclass correlation coefficient

ROC

Receiver operating characteristic

Grad-CAM

Gradient-weighted class activation mapping

k-NN

K-nearest neighbors

MLP

Multi-layer perceptron

Author Contribution

Yujian Wang: data curation, writing—original draft preparation, and investigation. Tongyu Wang: data curation and writing—original draft preparation. Fei Zheng: writing—original draft preparation. Wenhan Hao: Supervision. Qi Hao: visualization, investigation, and supervision. Wenjia Zhang: visualization, investigation, and supervision. Ping Yin: visualization, investigation, and supervision. Nan Hong: conceptualization, methodology, visualization, software, writing—reviewing and editing, and validation. All authors read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (No. 82471950); the Beijing Natural Science Foundation (Grant No. L242061).

Innovative Research Group Project of the National Natural Science Foundation of China,82471950,Nan Hong,Natural Science Foundation of Beijing Municipality,L242061,Nan Hong

Data Availability

The datasets generated or analyzed during the study are available from the corresponding author on reasonable request.

Declarations

Ethics Approval

This is an observational study. The Peking University People’s Hospital Research Ethics Committee has confirmed that no ethical approval is required. Written informed consent for participation is not provided by the participants’ legal guardians/next of kin because: as a retrospective study, it was approved by our institute committee without the informed consent of the patients.

Competing interests

The authors declare no competing interests.