Prediction of pathological risk subtypes of thymic anterior mediastinal cysts and thymic epithelial tumors based on CT radiomics and deep learning methods: a retrospective study

Abstract

Purpose statement

This study aims to develop a non-invasive model for preoperatively predicting the pathological risk classification of thymic anterior mediastinal cysts and thymic epithelial tumors using CT-based radiomics and deep learning. Accurate risk stratification before surgery can support personalized treatment planning and improve clinical outcomes.

Methods

A retrospective analysis was conducted on 144 patients with pathologically confirmed thymic anterior mediastinal cysts or thymic epithelial tumors who underwent preoperative thin-slice chest CT between January 2014 and December 2023. Regions of interest were manually segmented, and 1834 handcrafted radiomics features—including geometric, intensity, and texture features—were extracted using Pyradiomics. Deep learning features were derived from a ResNet50 network with transfer learning and cosine annealing learning rate adjustment. Radiomics and deep features were fused into a deep learning radiomics (DLR) feature set. Feature selection was performed before model training. The models were evaluated in training (n = 101) and test (n = 43) cohorts.

Results

The radiomics model achieved an AUC of 0.876 in the training cohort and 0.800 in the test cohort. The deep learning model yielded AUCs of 0.838 and 0.831, respectively. The combined DLR model showed superior performance, with an AUC of 0.964 in the training cohort and 0.820 in the test cohort, outperforming unimodal models in classification accuracy and robustness.

Conclusions

In this study, a model for predicting pathological risk classification of thymic anterior mediastinal cysts and thymic epithelial tumors was developed by combining radiomics and deep learning, and its superior prediction was confirmed in verification. The results show that the model is capable of preoperatively assessing the pathological risk classification of patients, which provides strong support for the need of individual treatment strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03169-z.

Thymic epithelial tumors (TETs) and thymic cysts (TCs) are the most common types of occupying lesions in the anterior mediastinum. The World Health Organization classifies the pathological staging of TETs into thymoma (A, AB, B1, B2, B3), thymic carcinoma (TCA), and neuroendocrine tumour [1]. Treatment options for TETs and TCs, as well as patient prognosis, differ significantly due to pathological staging [2], with some studies showing that patients with type A, AB or B1 tumors (low-risk thymomas) have a higher overall survival rate than patients with type B2 or B3 tumors (high-risk thymomas), and that thymic carcinomas have the worst prognosis, with a 5-year survival rate of 30–50% [3].The treatment of TETs and TCs is very different, asymptomatic TCs do not require surgical treatment clinically, whereas radical surgical resection is still the treatment of choice for TETs, and TETs with poor pathological typing and a high degree of malignancy can usually only be treated with palliative chemotherapy or radiotherapy [4, 5]. Currently, tissue biopsies are commonly used to obtain samples, which are combined with histological analyses and immunohistochemical staining for comprehensive evaluation to provide accurate pathological staging information [6]. However, tissue biopsy is invasive, inefficient, and relatively costly, and it is particularly difficult to obtain accurate preoperative pathological staging to guide clinical treatment in the anterior mediastinum due to its anterior proximity to the sternum, the numerous large blood vessels surrounding it, and the unavailability of tissues in patients with advanced or metastatic TETs [7]. Therefore, the development of a non-invasive and efficient method to predict pathological risk classification of anterior mediastinal masses is of great importance.

Over the past two decades, as a non-invasive test, CT has solved many problems in the clinical setting. And CT is an indispensable and important part in the diagnosis and management of anterior mediastinal masses. However, accurate diagnosis of anterior mediastinal masses by CT faces many challenges. First, the complex anatomical structures in the mediastinum make it difficult for the human eye to distinguish anterior mediastinal masses with similar density to normal tissues on CT, and it is not possible to accurately locate the anatomical position of anterior mediastinal masses [8]. Secondly, in the display of small structures or lesions, CT has a slightly lower resolution and obtains limited information, resulting in smaller lesions, especially early lesions, being easily missed. Thirdly, anterior mediastinal masses have a variety of pathological types, making it difficult for even imaging specialists who have been in practice for many years to distinguish specific pathological types by CT [9, 10]. Moreover, clinical studies have demonstrated a clear association between thymomas and myasthenia gravis, with approximately 30–50% of patients with thymoma exhibiting varying degrees of myasthenic symptoms [11]. Clinical management in such cases requires a comprehensive assessment that considers both pathological classification and neuromuscular involvement. However, relying solely on CT imaging is insufficient for fully evaluating the neuromuscular risk in these patients. To overcome these challenges, there is an urgent clinical need for a new technique to accurately diagnose anterior mediastinal masses.

With the rapid development of machine learning and artificial intelligence in the field of medicine, relevant researchers use radiomics to extract a large number of quantitative features from medical images, objectively quantify the digital features at the lesion, overcome the subjective limitations of human interpretation of medical images, assess the characteristics of the disease more objectively, establish diagnostic and prognostic models which assist doctors in better understanding and diagnosis of the disease [12]. Radiomics approaches have recently been used to predict histological subtype classification and pathological risk types of thymic epithelial tumors [13, 14]. However their studies are limited, small samples and a low number of quantitative features, model overfitting, poor generalisation and limited expressivity are inevitable. This study intends to develop a predictive model based on CT radiomics and deep learning, incorporating clinical factors, to improve the accuracy of preoperative diagnosis of anterior mediastinal cysts and thymic epithelial tumour pathology risk typing.

Methods

Study design and population

In this study, we constructed a radiomics imaging framework by novelly integrating radiomics data with deep learning methods. In addition, we combined deep learning-derived features with traditional omics features to build comprehensive deep learning radiomics features. Figure 1 shows the workflow of this study.

Fig. 1.

Workflow diagram of this study

In this study, we reviewed information on 308 patients with anterior mediastinal lesions at Tianjin Chest Hospital between 1 January 2014 and 31 December 2023 and included patients who met the following inclusion criteria: (1) underwent complete surgical resection of the anterior mediastinal mass at Tianjin Chest Hospital; (2) postoperative pathological confirmation of the lesion as mediastinal cysts or TETs. The exclusion criteria were as follows: (1) had a History of previous treatment; (2) Multiple lesions in the mediastinum; (3) Postoperative pathology that could not determine the type of pathology; (4) Missing, unavailable, or lack of preoperative thin-section CT documentation. After screening, 144 eligible patients were finally included in the study. After a 7:3 ratio randomisation split, 101 cases were used as the training cohort and 43 cases as the test cohort. Figure 2 shows the flow of inclusion and exclusion criteria for this study.

Fig. 2.

Flowchart of patient inclusion and exclusion criteria

Image acquisition and pre-processing

The latest preoperative chest CT scans were downloaded from Picture Archiving and Communication System (PACS), a picture archiving and communication system of Tianjin Chest Hospital, in the format of Digital Imaging and Communications in Medicine (DICOM). The CT images were preprocessed because the hospital uses different CT scanners (including PNMS, Siemens, and Philips), and there were differences in layer thickness, voxel size, window width, and window position between patients. We first resampled all the CT images and normalised the voxel units to 0.7 × 0.7 × 1.5. After that, we normalised the window width and window position to 350 and 40, respectively, which we found to be appropriate in the region of interest (ROI) segmentation process.

Segmentation of region of interest

The fully manual segmentation of the ROI (tumour) was performed by three experienced physicians: two thoracic surgeons, Z.W (with 11 years of experience in thoracic oncology) and D.Y (with 5 years of experience in thoracic oncology), and a radiologist, L.J (with 15 years of experience in medical imaging). Tumour contours were outlined in each of the three orthogonal planes and integrated by the software into a three-dimensional structure. Any disagreements that arose during the segmentation process were confirmed and guided by radiologist L.J. Thoracic surgeon S.D.Q (with 30 years of experience in cardiothoracic surgery) reviewed the segmentation results. ROI segmentation was carried out using the open source software ITK-SNAP (version 3.8.0).

Pathological risk assessment

Anterior mediastinal masses were classified according to the 2015 WHO histological classification [thymic cysts, thymoma types A, AB, B1, B2, B3 and thymic carcinoma (type C)] [1]. It has been shown that asymptomatic anterior mediastinal thymic cysts do not require surgical treatment clinically and have a better prognosis. Types A, AB and B1 thymomas are less invasive, and patients have a longer postoperative survival and a better prognosis. Types B2 and B3 thymomas are more invasive and have a worse prognosis than types A, AB and B1. Thymoma is the most aggressive and have the worst prognosis [15–17]. Therefore, we classified the anterior mediastinal masses into thymic cysts (TC); low-risk thymoma (LRT) including: types A, AB and B1; high-risk thymoma (HRT) including: types B2 and B3; and thymic carcinoma (TCA).

Imaging histological feature extraction

In this study, we classified the hand-crafted radiological features into three main categories: (I) geometric features; (II) intensity features; and (III) texture features. Geometric features aim to capture the 3D shape characteristics of the tumour. Intensity features use first-order statistical methods to analyse the statistical distribution of tumour voxel intensities. In contrast, texture features assess the voxel intensity patterns and spatial distribution through more sophisticated second-order and higher-order analyses. We used a variety of texture feature extraction techniques, including grey-level co-occurrence matrix (GLCM), grey-level run-length matrix (GLRLM), grey-level size-zone matrix (GLSZM) and neighbourhood grey-level difference matrix (NGTDM).

Imaging genomics modelling

In the feature selection phase of the study, we used a multi-step approach. Firstly, we standardised the features using Z-values, supplemented by p-value calculations based on t-tests, and chose to retain features with p-values below 0.05. Subsequently, we scrutinised the repeatable features using the Pearson correlation coefficient and chose to retain only one feature when the correlation of the feature pairs exceeded 0.9, thus reducing redundancy using a greedy recursive deletion strategy. Finally, to further refine the feature set of radiometric features, we employed LASSO regression to effectively reduce the importance of irrelevant features. The optimal regularisation parameter λ for this process was determined through 10-fold cross-validation, and after Lasso-based feature refinement, we assessed the risk using a linear model (LR) model, as well as a deep learning-based Multilayer Perceptron (MLP) model. For model hyperparameter tuning, we performed 5-fold cross-validation on the training set and hyperparameter optimisation using the Gridsearch algorithm. The model parameters with the best median performance were selected for the final model training.

Deep learning models

In our image processing for deep learning, we selected slices that show the largest region of interest (ROI) for each patient as representative images. To reduce the complexity of the algorithmic analysis and to reduce background noise, we retained only the smallest bounding rectangle containing the ROI. This rectangle was enlarged by a further 10 pixels, the decision based on recent studies that have highlighted the importance of the peritumoural region. We normalised the intensity distribution across the RGB channels by Z-score normalising the images. These normalised images were then used as input to the model. During the training phase, we implemented a real-time data enhancement strategy including random cropping, horizontal flipping and vertical flipping. For the test images, we limited ourselves to normalisation. In this study, we used the well-known neural network ResNet50 to extract deep learning features. In our study, to ensure the effectiveness of the model in various patient populations with significant differences, we implemented transfer learning. This process involves initialising the model with pre-trained weights from the ImageNet database, thus enhancing its adaptability to different datasets. A key aspect of our approach is the careful tuning of the learning rate to facilitate better generalisation across different datasets. To this end, we employ a cosine decay learning rate strategy, which is defined as follows:

$${\eta }_t={\eta }_{\text{min}}^i+\frac{1}{2}\left({\eta }_{\text{max}}^i,-,{\eta }_{\text{min}}^i\right)\left(1,+,c,o,s,\left(\frac{T_{\mathit{cur}}}{T_i},\pi \right)\right)$$

${\eta }_{\text{min}}^i=0$ Denotes the minimum learning rate, ${\eta }_{\text{max}}^i=0.01$ denotes the maximum learning rate. $T_i=30$ Denotes the number of calendar elements in the iterative training process. Other basic hyperparameters include the use of the stochastic gradient descent (SGD) as the optimiser and the use of the soft-maximum cross entropy as the loss function.

In our model, the output probabilities computed by the CNN are defined as deep learning features.

Deep learning of imaging histological features

For the best deep learning model in our study, the CNN model showed the best performance in the test set. We utilised its penultimate layer for feature extraction. Given the complexity of the CNN model, to reduce the risk of overfitting associated with this high dimensionality, we used principal component analysis (PCA) to reduce these features to a more manageable 64 dimensions. To construct deep learning radiomics (DLR) features, we used a pre-fusion algorithm that combines 64-dimensional deep learning features with 1834-dimensional radiomics features. This process results in a total of 1898-dimensional features. Subsequently, we used a similar procedure to radiomics for feature selection and model construction. To build a robust clinical model, we performed univariate analysis of clinical features using the same methodology as for radiomics and deep learning radiomics (DLR) signatures. This approach helps to combine clinical insights with imaging and algorithmic evaluation. We evaluated the discriminative performance of all models across the three classification types using Micro and Macro AUC metrics to assess the effectiveness of our algorithmic approach. In addition, we performed individual versus other AUC analyses for each category to further determine the ability of the models to discriminate between categories.

Statistical analyses

We randomly divided the dataset, assigning 70% to the training group and the remaining 30% to the internal test group. The results (shown in Table 1) shows no statistically significant difference (p-value greater than 0.05) between the training and test samples, ensuring a fair division of the data. We performed the analyses using Python version 3.7.12 and statsmodels version 0.13.2. Machine learning models were developed using the scikit-learn version 1.0.2 interface. For deep learning training, we used NVIDIA 4090 GPUs, and the MONAI 0.8.1 and PyTorch 1.8.1 frameworks.

Table 1.

Clinical baseline characteristics of patients included in the study

	ALL		Train cohort	Test cohort	P value
Age (mean ± SD)		52.70 ± 12.77	52.05 ± 13.20	54.23 ± 11.69	0.478
Gender					0.636
Female (%)		73(50.69)	53(52.48)	20(46.51)
Male (%)		71(49.31)	48(47.52)	23(53.49)
Pathological type
Thymic cyst (%)		13(9.03)	10(9.90)	3(6.98)
Low-risk thymomas (%)		66(45.83)	44(43.56)	22(51.16)
High-risk thymomas (%)		36(25.00)	25(24.75)	11(25.58)
Thymic carcinomas (%)		29(20.14)	22(21.78)	7(16.28)
Myasthenia_Gravis					1.0
No (%)		141(97.92)	99(98.02)	42(97.67)
Yes (%)		3(2.08)	2(1.98)	1(2.33)

Results

Patient clinical characteristics

The baseline characteristics of the 144 patients are shown in Table 1.In addition to the outcome metrics, we compared the overall distributions of the training and test cohorts for this randomly grouped scenario. The results show that there were no significant differences in any of the information between the two cohorts. In the training cohort, 52.48% of the patients were female and 47.52% were male, 9.90% were pathologically typed as thymic cysts, 43.56% as low-risk thymomas, 24.75% as high-risk thymomas, and 21.78% as thymic carcinomas, and 98.02% were myasthenia gravis-negative and 1.98% were myasthenia gravis-positive. In the test cohort, 46.51%of patients were female and 53.49% were male, 6.98% were pathologically staged as thymic cysts, 51.16% as low-risk thymomas, 25.58% as high-risk thymomas, and 16.28% as thymic carcinomas, and 97.67% were myasthenia gravis-negative and 2.33% were myasthenia gravis-positive.

Feature extraction and screening

For the ROIs manually outlined on the CT images (Fig. 3), a total of 1834 features were extracted using PyRadiomics. The Z-value was used to standardise the features, supplemented by the p-value calculation based on the t-test, and the remaining 279 features were selected to retain those with p-values lower than 0.05. Ninety-two features were retained after excluding features by Spearman correlation test. For maximum interface ROI, 63 simplified deep learning features were extracted using ResNet50 convolutional neural network. Subsequently, we did post-fusion splicing of the imageomics features with the deep learning features, and constructed 20 features related to prediction by LASSO regression and multivariate COX regression screening (Supplementary Figs. 1 and 2). The details of the extracted features are shown in Supplementary Tables 1–3.

Fig. 3.

Region of interest segmentation and feature extraction. A Region of interest localization on CT images. B Manual segmentation of tumors on CT images. C–D extracting the maximum cross-section of ROIs, respectively, and extracting deep learning features using convolutional neural networks

Training and testing of the model

The performance of the three different models was evaluated in the training cohort and test cohort, respectively (Fig. 4 and Table 2). The integration of deep learning with omics features in Deep Learning Radiomics Signature shows a significant improvement in model performance, with an AUC of 0.964 (95% confidence interval (CI) 0.947–0.981) for Micro in the training cohort of the Deep Learning Radiomics model, and an AUC of 0.820 for Micro in the test cohort (95%. CI 0.749–0.891). Although performance in the test cohort decreased slightly, these models maintained strong discriminative power. The model with the highest AUC in the training cohort was TC in DLRad_model with an AUC value of 0.983 (95%, CI 0.965–1.000). The model with the highest AUC in the test cohort was TC in DL_model, with an AUC value of 0.885 (95%, CI 0.765–1.000) demonstrating excellent performance, with metrics reflecting not only high accuracy but also a balance of sensitivity and specificity that ensures reliable classification in different clinical settings. Deep learning radiomics models were the best performing in all cohorts, and this fusion approach leverages the advantages of deep learning and holographic data to significantly improve diagnostic accuracy and model robustness compared to traditional radiomics or standalone deep learning models, highlighting the potential of multimodal strategies in advancing precision medicine.

Fig. 4.

ROC curves of different models in the training and test cohorts respectively. A ROC curves of radiomics model in the training and test cohorts respectively. B ROC curves of deep learning model in the training and test cohorts respectively. C ROC curves of radiomics deep learning model in the training and test cohorts respectively

Table 2.

Performance evaluation of the different models in the training and test cohorts

	Acc	AUC(95%CI)	Sensitivity	Specificity	PPV	NPV
Train cohort
Rad_model
Micro	0.824	0.876(0.839–0.914)	0.337	0.987	0.895	0.817
TC	0.832	0.901(0.842–0.960)	0.727	0.912	0.865	0.812
LRT	0.901	0.871(0.759–0.984)	0.600	0.934	0.500	0.955
HRT	0.733	0.795(0.703–0.886)	0.720	0.737	0.474	0.889
TCA	0.703	0.881(0.811–0.952)	0.909	0.646	0.417	0.962
DL_model
Micro	0.807	0.838(0.796–0.881)	0.396	0.944	0.702	0.824
TC	0.782	0.876(0.811–0.941)	0.750	0.807	0.750	0.807
LRT	0.842	0.856(0.743–0.969)	0.700	0.857	0.350	0.963
HRT	0.772	0.808(0.713–0.904)	0.600	0.829	0.536	0.863
TCA	0.634	0.731(0.619–0.843)	0.818	0.582	0.353	0.920
DLRad_model
Micro	0.901	0.964(0.947–0.981)	0.703	0.967	0.877	0.907
TC	0.921	0.983(0.965–1.000)	0.886	0.947	0.929	0.915
LRT	0.960	0.977(0.935–1.000)	0.800	0.978	0.800	0.978
HRT	0.822	0.913(0.856–0.970)	0.920	0.789	0.590	0.968
TCA	0.881	0.953(0.910–0.996)	0.909	0.873	0.667	0.972
Test cohort
Rad_model
Micro	0.779	0.800(0.726–0.874)	0.163	0.984	0.778	0.779
TC	0.791	0.825(0.695–0.954)	0.864	0.714	0.760	0.833
LRT	0.953	0.783(0.356–1.000)	0.333	1.000	1.000	0.952
HRT	0.674	0.739(0.572–0.905)	0.636	0.687	0.412	0.846
TCA	0.651	0.734(0.519–0.949)	0.714	0.639	0.278	0.920
DL_model
Micro	0.837	0.831(0.755–0.908)	0.512	0.946	0.759	0.853
TC	0.860	0.885(0.765–1.000)	0.864	0.857	0.864	0.857
LRT	0.605	0.583(0.213–0.953)	0.333	0.625	0.062	0.926
HRT	0.791	0.778(0.622–0.935)	0.545	0.875	0.600	0.848
TCA	0.488	0.694(0.506–0.883)	0.857	0.417	0.222	0.937
DLRad_model
Micro	0.797	0.820(0.749–0.891)	0.558	0.876	0.600	0.856
TC	0.767	0.866(0.761–0.971)	0.727	0.810	0.800	0.739
LRT	0.930	0.758(0.306–1.000)	0.333	0.975	0.500	0.951
HRT	0.535	0.665(0.495–0.835)	0.818	0.437	0.333	0.875
TCA	0.814	0.714(0.479–0.949)	0.429	0.889	0.429	0.889

Grad-CAM visualisation

To investigate the recognition ability of deep learning models on different samples, we used the Gradient-weighted Class Activation Mapping (Grad-CAM) technique for visualisation Figure 5. Illustrates the application of Grad-CAM, highlighting the activations in the final convolutional layer that are relevant to the prediction of cancer type, and these visualisations demonstrate well how the model makes predictions based on different regions of the image. This helps to identify regions of the image that have a significant impact on the model's decisions, providing insight into the interpretability of the model.

Fig. 5.

Grad-CAM visualizations for 2 typical samples (‘136’ and ‘155’ respectively)

Discussion

Thymic epithelial tumors (TETs) and thymic cysts (TCs) are the most common types of occupying lesions in the anterior mediastinum. Due to differences in pathological staging, there are significant differences in treatment options for TETs and TCs as well as patient prognosis [2]. Therefore, it is necessary to be able to make preoperative predictions of patients' anterior mediastinal masses with accurate pathological risk typing. Currently, tissue biopsies combined with histological and immunohistochemical analyses provide accurate pathological staging information in clinical practice. However, tissue biopsy is invasive, inefficient, and costly, and the anterior mediastinum increases the difficulty of the procedure due to its proximity to the sternum and large vessels. In addition, tissue samples are difficult to obtain in patients with advanced or metastatic TETs, and it is particularly difficult to obtain accurate pathological typing preoperatively to guide clinical treatment [7]. Therefore, it is necessary to seek a new method to preoperatively assess the pathological risk typing of TETs and TCs. In this study, we used CT image-based histology and deep learning methods to predict pathological risk subtypes of thymic anterior mediastinal cysts and thymic epithelial tumors. In both the training and test groups, the deep learning combined with radiomics model performed better, with larger AUC values and higher accuracy than the unimodal radiomics model and deep learning model. This fusion approach takes full advantage of deep learning and holographic data to significantly improve diagnostic accuracy and model robustness, highlighting the potential of multimodal strategies in advancing precision medicine.

Over the past 2 decades, CT has been widely used in the preoperative diagnosis of TETs and TCs, and the typical CT presentation of TC is a low-density rounded mass with clear boundaries and homogeneous density; LRT is a low-density rounded mass with clear boundaries and homogeneous or slightly inhomogeneous density; HRT is an irregular mass with unclear boundaries, inhomogeneous density, and necrotic cystic degeneration. TCA presents as an irregular mass with poorly defined borders, often invading the surrounding tissue, with uneven density and often necrotic cystic changes. However, when TC and LRT are complicated by inflammation or haemorrhage, it is difficult to distinguish them from HRT or TCA on CT, which makes subsequent treatment difficult. One study shows that the CT diagnostic accuracy of TETs was 90.1%, while that of TC was only 42.3%, with 80.5% of TC patients being misdiagnosed as TETs [18]. Radiomics features have been widely used for preoperative prediction of TETs and TC [14, 19]. Zhang et al. [20], developed CT-based radiomics maps integrating rad scores and conventional CT manifestations as a valid tool to differentiate between TCs and TETs. In their test cohort, the AUC value was 0.953 (0.893–0.997), sensitivity was 0.917 (0.715–0.985), specificity was 0.867 (0.584–0.977), and accuracy was 0.897, giving an overall excellent model performance. However, their study was limited to distinguishing between TCs and TETs and did not further classify the risk typing of TETs. In this study, we not only distinguished between TCs and TETs, but also accurately predicted LRT, HRT, and TC in TETs.

ResNet50 is a classical convolutional neural network (CNN) widely used in medical image recognition and semantic segmentation. It contains 50 convolutional layers, and by adding ‘jump connections’ in each residual module, it achieves direct transfer of information and avoids the disappearance of features layer by layer, so as to improve the expressiveness and accuracy of the network [21, 22]. Migration learning makes use of models trained on one task and applies them to another related task. With migration learning, the knowledge and features of a pre-trained model can be efficiently migrated to a new task, reducing annotation data requirements, accelerating training and improving performance.ResNet50 significantly improves the performance of a medical image recognition task after pre-training on a large-scale dataset (e.g., ImageNet) [23, 24]. In this study, we used the well-known neural network ResNet50 to extract deep learning features, and to ensure the effectiveness of the model in various patient populations with significant differences, we implemented migration learning. Additionally we fine-tuned the learning rate to allow for better generalisation across datasets. The final results show significantly better classification of TC, LRT, HRT, and TCA in the test cohort deep learning model than in the radiomics model, and this advantage demonstrates the potential of deep learning models to revolutionise diagnostic methods, providing more detailed and accurate analyses than traditional radiological histology methods. These results affirm the critical role that advanced deep learning architectures can play in improving precision medicine. In addition, we post-fused 92 features extracted and filtered from radiomics with 63 simplified deep learning features extracted using ResNet50 convolutional neural network, and constructed a deep learning radiomics model by filtering 20 features related to prediction. The model efficacy was the top performer in all cohorts, and this fusion approach took full advantage of deep learning and holographic data to significantly improve diagnostic accuracy and model robustness compared to traditional radiomics or stand-alone deep learning models, highlighting the potential of multimodal strategies in advancing precision medicine.

However, our study still has some limitations. Firstly, this is a retrospective study and may suffer from selection bias. Second, our model performs erratically in Sensitivity, which we suspect is low due to the small sample size included resulting in the model not being able to adequately learn more diverse features of the data during training leading to lower Sensitivity, and in the case of small samples, overfitting is prone to occur in the model, which severely affects the performance of the model. These performances reflect that our sample size needs to be further expanded. Third, as we are a single-centre study, the demographics and ethnicity of the sample are limited, which may limit the external validity and generalisability of our findings across different populations, and more robust external validation needs to be achieved using multi-centre samples in the future. Fourthly, radiomics features extracted by manual segmentation are subjective and variable, and time-consuming and labour-costly, and in our next work we will mainly focus on developing automated segmentation to improve efficiency and consistency and reduce subjectivity. Therefore, future larger, multicentre, prospective studies are essential to validate our findings.

Conclusion

In a word, this study combined radiomics and deep learning to develop a model for predicting pathological risk subtypes of anterior mediastinal cysts and thymic epithelial tumors of the thymus and confirmed the superior predictive efficacy of the model in a test set. The results suggest that the model is capable of preoperatively assessing pathological risk subtypes in patients with anterior mediastinal cysts and thymic epithelial tumors, thereby providing strong support for the development of individualised treatment strategies for patients.

Abbreviation

TETs

Thymic epithelial tumors

TCs

Thymic cysts

TCA

Thymic carcinoma

ROI

Region of interest

CNN

Convolutional neural network

DLR

Deep learning radiomics

AUC

Area under the curve

PPV

Positive predictive value

NPV

Negative predictive value

DL_model

Deep learning_model

Rad_model

Radiomics_model

DLRad_model

Deep learning radiomics_model

Author contributions

WZ: investigation, methodology, data curation, writing original draft, review & editing; XZ: investigation, methodology, formal analysis, writing, review & editing; YD: investigation, methodology, formal analysis, software; QL: investigation, methodology; jl: supervision, validation; DS: conceptualization, project administration, writing, review & editing. All authors read and approved the final manuscript.

Funding

Tianjin Key Medical Discipline (Thoracic Surgery) Construction project (TJYXZDXK-018A).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. This study was approved by the ethical review committee of Tianjin Chest Hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.