Abstract

Introduction:

Cervical cancer remains a major global health burden, and accurate pathological classification is essential for personalized treatment planning. However, conventional radiomics studies often rely on manual lesion delineation and are limited in extracting meaningful imaging biomarkers from heterogeneous cervical cancer lesions.

Methods:

We proposed a convolutional recurrent feature extraction (CRFE)-based automatic segmentation framework for cervical cancer MRI images and developed histogram-based imaging features reflecting lesion pixel concentration trends. These features were integrated with conventional radiomics and clinical features. Feature engineering and machine learning classifiers, including random forest (RF), XGBoost, support vector machine, and logistic regression, were evaluated to construct an auxiliary diagnostic model for pathological classification. The dataset included 114 patients with cervical cancer who underwent MRI examinations.

Results:

The CRFE segmentation model achieved an Intersection over Union (IoU) of 0.9443, a Dice coefficient of 0.5980, and an F1-score of 0.7085. Feature selection retained 30 key imaging biomarkers, including the median of the histogram, GLSZM large-area low gray-level emphasis (LoG, σ = 2.0mm, 3D), and GLRLM long-run low gray-level emphasis (LoG, σ = 2.0mm, 3D). Among the evaluated classifiers, the RF model achieved the best performance, with an accuracy of 87.27% and an F1-score of 86.91% in pathological classification.

Discussion:

The proposed deep learning–radiomics framework enables accurate lesion segmentation and effective pathological classification of cervical cancer. This auxiliary diagnostic model may reduce unnecessary invasive procedures and improve early screening and clinical decision-making.

1 Introduction

Radiomics, an emerging medical image analysis technology, objectively describes the shape, texture, density, and spatial distribution of tumors (

Mayerhoefer et al., 2020

;

Lambin et al., 2017

) based on high-throughput quantitative features extracted from computed tomography (CT) (

Berenguer et al., 2018

), magnetic resonance imaging (MRI), or positron emission computed tomography (PET) (

Cook et al., 2014

) images and provides a detailed description of the phenotypic information of tumors (

Gillies et al., 2016

). For cervical cancer, MRI radiomics can effectively distinguish between cervical squamous cell carcinoma and adenocarcinoma and predict lymph-node metastasis (

Follen et al., 2003

). However, in existing radiomics research of cervical cancer, the delineation of the lesion area heavily relies on manual labeling (

Oussahi et al., 2024

), which cannot effectively analyze large-scale image data (

Zabihollahy et al., 2022

). Furthermore, the regional heterogeneity of cervical cancer lesions is high, and the image features are difficult to reliably extract (

Kumar et al., 2012

). This hinders the construction of effective biomarkers and makes rapid diagnosis and early screening difficult (

Shah et al., 2022

;

Kessler, 2017

). In order to solve this problem, we first designed a deep learning-based lesion area extraction model for cervical cancer. Specifically, based on the U-Net architecture (

Zhou et al., 2018

), we combined convolutional networks, a recurrent structure, and an attention mechanism to form the convolutional recurrent feature extraction (CRFE) model, which improves contextual modeling and key-region recognition, thereby enabling automatic segmentation of cervical cancer lesions. On this basis, we designed a variety of imaging features for the lesion area, including gray-level run length matrix–long run low gray-level emphasis (GLRLM–LRLGLE), gray-level size zone matrix–large area high gray-level (GLSZM–LAHGL), and median of histogram, and added these features to existing radiomics features. Combined with clinical features, feature screening was performed to identify biomarkers for the diagnosis of cervical cancer. Finally, an auxiliary diagnosis model of cervical cancer pathological classification based on the biomarkers of the imaging group was formed. The main contributions of this study are summarized as follows:

We designed a deep learning lesion region extraction method for cervical cancer images. Based on the U-Net architecture (Zhou et al., 2018), our method integrates convolutional networks, cyclic structures, and attention mechanisms (Niu et al., 2021; Brauwers and Frasincar, 2021) and uses Dice loss and cross-entropy loss to construct the CRFE model. The Intersection over Union (IoU) of the model is 0.9443, the Dice coefficient is 0.5980, and the F1-score is 0.7085, which can accurately extract the cervical cancer lesion area.

Based on characteristics of cervical cancer lesion area images, we designed a series of imaging features, including the median of histogram, to perform more comprehensive radiomics feature extraction.

We added designed features to the existing radiomics features and integrated the clinical features to construct an auxiliary diagnosis model of cervical cancer pathological classification. The results show that the median of the histogram, GLSZM large area low gray-level emphasis (LoG, , 3D), GLRLM long run low gray-level emphasis (LoG, , 3D), gray-level dependence matrix (GLDM) dependence variance (LoG, , 3D), and 27 other features have a better ability to predict cervical cancer lesions than clinical features alone. This may help reduce unnecessary clinical examinations and improve the efficiency of early screening and clinical diagnosis of cervical cancer.

2 Materials and methods2.1 Dataset

This study is a single-center retrospective study (Talari and Goyal, 2020) and was approved by the Ethics Committee of the First Hospital of Lanzhou University (with informed consent waived due to anonymous data analysis). A total of 114 patients with cervical cancer who were treated in our hospital from January 2022 to December 2024 were included in the study. All subjects underwent MRI examination before the initial treatment.

The inclusion criteria included the following:

Patients with cervical cancer diagnosed by pathology as squamous cell carcinoma or adenocarcinoma.

FIGO (International Federation of Gynecology and Obstetrics) staging (Bhatla et al., 2019) was IA to IIIC.

Newly diagnosed patients without surgical treatment.

The age ranged from 28 to 80 years old, with a Karnofsky (Péus et al., 2013) score of points.

Patients who underwent standard MRI acquisitions, including T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC).

FIGO staging was used to classify the tumor extent. Stage IA represents microscopic lesions that are confined to the cervix, whereas stage IIIC indicates pelvic and/or para-aortic lymph-node metastasis. Stage IIIC disease implies more widespread cancer involvement. Because nodal metastasis can alter the tissue contrast on MRI, stage IIIC lesions may be more distinguishable in the proposed model. The Karnofsky Performance Status (KPS) score assesses a patient’s functional ability, ranging from 0 to 100; with scores 70 indicating that the patient can care for themselves.

The exclusion criteria included the following:

Patients who have undergone surgery / other systemic treatment.

Patients with poor MRI image quality / missing key sequences.

Patients with severe infectious diseases / autoimmune diseases / blood diseases / liver diseases / other malignant tumors.

Patients with extensive metastatic disease beyond the pelvic region.

The median age of the included patients was 51 years (range: 28–80 years). All images were acquired using the same MRI system (Siemens 3.0-T MAGNETOM Skyra and Philips Ingenia 3.0-T). Image quality control consisted of standardized slice thickness, field-of-view consistency, signal-to-noise ratio assessment, and the exclusion of studies with motion artifacts (Table 1).

CharacteristicCategory/StatisticValue (n = 114)Percentage95% Confidence intervalAgeMean SD51.2 8.7​(49.6, 52.8)Median (range)51 (28–80)​​HPV statusPositive7868.40%(59.2%, 76.8%)Negative97.90%(3.7%, 14.5%)Not reported2723.70%(16.2%, 32.7%)Lymph node metastasisAbsent9583.30%(75.3%, 89.6%)Present1916.70%(10.4%, 24.7%)Lymphovascular invasionAbsent10289.50%(82.3%, 94.4%)Present1210.50%(5.6%, 17.7%)Depth of cervical invasion (cm)Mean SD0.54 0.28​(0.49, 0.59)Median (range)0.5 (0.1–1.0)​​FIGO stageStage I (IA/IB)4842.10%(32.9%, 51.8%)Stage II (IIA/IIB)5144.70%(35.5%, 54.3%)Stage III/IV1210.50%(5.6%, 17.7%)Unknown32.60%(0.6%, 7.5%)TreatmentChemotherapy (yes)8877.20%(68.5%, 84.6%)Radiotherapy (yes)5850.90%(41.4%, 60.3%)

Summary of the baseline characteristics, clinicopathological features, and treatment methods of the patients (n = 114). The table shows the age, HPV infection status, lymph-node metastasis, lymphatic vascular invasion, cervical infiltration depth, clinical stage, and chemotherapy and radiotherapy of 114 patients included in the study. The data are presented as the mean standard deviation, median (range), frequency (percentage), and 95% confidence interval. Patients with unknown FIGO stage were included to avoid unnecessary exclusion and potential selection bias. Treatment information was recorded as baseline clinical data and was not used as a predictive variable in model training.

2.2 CRFE segmentation model

In order to realize the automatic and accurate segmentation of the lesion area in cervical cancer images, this study proposes a segmentation model based on the CRFE architecture. The model utilizes the traditional convolutional neural network (CNN) (Zhao et al., 2024) to extract spatial features, combines the loop structure to capture the contextual connections (Chen et al., 2016; Zuo et al., 2016) in the image, and strengthens the key region features through the attention mechanism. Strengthening key regional features enhances the representation of clinically relevant areas, whereas biasing the model toward specific key features may reduce generalizability. The attention mechanism was introduced to adaptively weight important lesion regions without overemphasizing noise, thereby improving both interpretability and segmentation accuracy. The implementation process of the model includes three aspects: data preprocessing, the CRFE framework, and the training strategy.

2.2.1 Data preprocessing

In order to ensure a unified input data format and enhance the generalization ability of the model, we preprocessed the original MRI image data of 141 patients with cervical cancer from the Department of Gynecology, First Clinical Medical College of Lanzhou University (114 patients had complete clinical and imaging data). MRI images were obtained using two different 3.0-T magnetic resonance scanners (Siemens MAGNETOM Skyra and Philips Ingenia). Analysis utilized three sequences, including T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and their corresponding apparent diffusion coefficient (ADC) maps. First, all images are resampled to a uniform resolution to ensure the consistency of the input images on the spatial scale (Anthony Parker et al., 2007; Yu, 2002). Subsequently, the intensity standardization is carried out, and the z-score standardization (Curtis et al., 2016; Gao et al., 2019) is applied to the image to eliminate the difference in grayscale distribution caused by different acquisition equipment and scanning parameters (Saravanan, 2010; Lissner et al., 2012). In order to expand sample diversity and improve robustness (Kitano, 2004; Gowal et al., 2021), a random data enhancement strategy was adopted. It includes random rotation of the image in the range of 15°, random flipping of the image in the horizontal and vertical directions (Zhong et al., 2020; Takahashi et al., 2019), random scaling of the image in the range of 0.9–1.1 times, adding Gaussian noise (Gedraite and Hadad, 2011; Ahmad B. et al., 2022; Pyatykh et al., 2012) with a standard deviation of 0.01 to improve the robustness of the model to the noise, and dividing the dataset by 70% (99 cases) of the training set, 15% (21 cases) of the validation set, and 15% (21 cases) of the test set.

2.2.2 CRFE framework

The CRFE segmentation framework was constructed by adopting the encoder–recursive feature extraction module–decoder structure. The encoder part is responsible for extracting the multi-scale spatial features of the input MRI image, and the input layer accepts a two-dimensional image with a size of 256 256. The three convolution blocks adopt 32 3 3 convolution kernels, 64 3 3 convolution kernels, and 128 3 3 convolution kernels, respectively, and cooperate with the ReLU activation function to extract image features from low to high (Schmidt-Hieber, 2020). By using a 2 2 pooling window, downsampling is carried out (Wu et al., 2023), the feature map size is reduced, and the feature invariance is enhanced. To incorporate global context information, the module also enhances feature extraction by introducing a cyclic structure. Based on the features extracted by convolutional coding, 256 long short-term memory (LSTM) units are set up to learn the local and global spatio-temporal dependencies in the image. In addition, the module captures both forward and reverse information to fully model the complex context dependencies that may exist in the image. Notably, we introduce the attention module after the loop layer and assign adaptive weights according to the importance of different regions to highlight the features of key regions and alleviate the problem of feature redundancy (Ma et al., 2019). The spatial attention module can establish a rich context model on local features and encode wider context information into local features, thereby enhancing its representation ability.

Given the local feature (where denotes the input local convolutional feature map and , , and represent the number of channels and the spatial height and width, respectively), it is first passed through a convolutional layer to generate two new feature maps and , each of size . These feature maps are then reshaped to , where denotes the total number of pixels. This reshaping converts the spatial dimensions into a single dimension, enabling pixel-wise computation in the subsequent attention operation. After reshaping, matrix multiplication is performed between and , and the result is normalized using a softmax layer to produce the spatial attention map (Hu et al., 2018).

In Equation 2, is the influence of position i on position j. A more similar feature representation of the two locations contributes to a greater correlation between them. In parallel, the original feature map is passed through an additional convolutional layer to generate another feature map , which is likewise reshaped into to match the dimensionality required for the attention computation. Subsequently, matrix multiplication is performed between and the spatial attention map obtained from Equations 1, 2, and the resulting feature representation is reshaped back to . Finally, as described in Equation 3, the output feature is obtained by multiplying the aggregated context feature by the learnable coefficient and adding it element-wise to the original feature , producing the following attention-enhanced feature:

Here, denotes the output feature vector at position , which is enhanced by global contextual information; denotes the attention weight indicating the contribution of position to ; denotes the context feature at position ; and denotes the original local feature at position , which is used to establish a residual connection. Here, is initialized to 0, and more weights are gradually assigned, and it can be inferred that the result feature E at each position is the weighted sum of the features at all the positions and the original features. Therefore, by incorporating the global context view and selectively aggregating the features according to the spatial attention map, similar semantic features are enhanced, thereby improving intra-class compactness and semantic consistency. In the decoder part, up-sampling is used to enlarge the low-resolution feature map generated by the encoder. The up-sampled feature map is refined using 64 convolution kernels, and 32 convolution kernels are used to further fuse local information. The convolution kernel is used to generate a single-channel segmentation probability map, and the final segmentation result is output by the sigmoid activation function, as shown in Equation 4 (Narayan, 1997). The original image size is gradually restored by up-sampling and convolution operations, and the final segmentation result is generated.where represents the sigmoid activation function, denotes the input value to the activation unit, and is Euler’s number, which is the base of the natural logarithm. The sigmoid function maps any real-valued input into the range (0, 1).

The framework fully integrates the information from convolution features and recurrent modules based on MRI images and highlights key features using an attention mechanism, effectively improving the segmentation accuracy and robustness of cervical cancer lesions. The model framework is shown in Figure 1.

Convolutional recurrent feature extraction framework. The input is a MRI image, and after data preprocessing (resampling, Z-score standardization, and data enhancement), it enters the encoder (including convolutional layer Conv 128/64/32-ch, maximum pooling down-sampling, and skip connection). Skip connections refer to feature map connections between corresponding encoder and decoder layers to preserve spatial information. The feature then captures spatio-temporal dependencies and fuses bi-directional context information through a module containing bi-directional LSTM (256 units) and then adaptively weights key lesion-related features through the attention mechanism to suppress redundant features. LSTM denotes a long short-term memory network used to capture contextual dependencies in the feature maps. The decoder part includes upsampling , convolution layer (Conv 64/32-ch, and Conv + sigmoid), and finally the output ROI segmentation map of the same size . The independent CRFE module is also marked in the figure.

2.2.3 Model training strategy

In order to ensure that the model converges quickly and stably in the segmentation task, a series of optimized training strategies were adopted in this study. First, in the design of loss function, Dice loss (Zhao et al., 2020) and cross-entropy loss (Zhang and Sabuncu, 2018) are used to balance the accuracy of region overlap and pixel classification. Specifically, the Dice loss is defined as shown in Equation 5:where denotes the number of pixels correctly predicted as the lesion region, and represent the total number of predicted and ground-truth positive pixels, respectively, and is a small constant added for numerical stability.

The cross-entropy loss is defined as shown in Equation 6:where is the ground-truth binary label, is the predicted probability for the corresponding class, and denotes the natural logarithm. This loss penalizes incorrect high-confidence predictions.

The final combined loss function is defined in Equation 7:Here, represents the number of sub-modules (such as encoder/decoder), and is the binary segmentation loss (Dice/cross entropy) of each module, which supervises local feature learning. represents the number of global constraint terms, and is weighted by the hyperparameter to ensure the spatial consistency of prediction. is based on the prior clinical size of cervical cancer lesions (average diameter 1 cm–4 cm) (Horn et al., 2007), and any prediction that deviates from the medically reasonable range is penalized. This design forces the model to distinguish the lesion (positive class) from the surrounding tissue (background).

For the selection of the optimizer, this study used the Adam optimizer (Zhang, 2018), with an initial learning rate of 0.001. In order to ensure smooth parameter updates during training, the cosine-annealing strategy (Zhang et al., 2024) was used to gradually decay the learning rate. The cosine-annealing learning rate schedule is defined in Equation 8:Here, denotes the learning rate used at iteration . The symbols and represent the minimum and maximum learning rates within the th cosine annealing cycle, respectively. The variable refers to the current number of iterations completed in the ongoing cycle, while denotes the total number of iterations assigned to the th cycle. The cosine function is utilized to generate a smooth learning rate decay curve, and denotes the mathematical constant pi. Together, these components construct a cosine-annealing schedule that gradually reduces the learning rate from to within each cycle, thereby facilitating stable convergence and improving the robustness of the training process.

During the training process, the batch size is set to 4, which not only ensures the rational use of video memory but also ensures the stability of training. The number of training rounds (epochs) is set to 100 to ensure that the model can fully converge. In addition, in order to prevent over-fitting, an early stopping strategy based on the Dice coefficient of the validation set was adopted. Specifically, the monitoring index is the Dice coefficient on the validation set. When the improvement of the Dice coefficient in 10 consecutive epochs is less than 0.001, the training will stop in advance, thus avoiding the occurrence of overfitting (Table 2).

ComponentConfiguration detailsLoss functionOptimizerAdam (lr = 0.001)Learning rate adjustmentsCosine-annealing strategyBatch size4Epochs100Early stop strategyThere was no improvement in the Dice coefficient of the validation set in 10 rounds

Training strategy. lr, learning rate.

2.3 Biomarker construction and radiomics analysis2.3.1 Data preparation and processing

In this study, a multi-source fusion dataset was constructed, which integrated three types of complementary feature information: first, clinical data obtained from the clinical process, including patients’ clinical information and pathological test indicators; second, histogram features extracted from medical images that reflect the gray distribution of the image; third, high-dimensional radiomics features extracted by radiomics analysis technology, covering multi-level image features such as texture, shape, and transformation. All data are merged with the patient’s unique identification as the matching key to ensure the consistency of the model data.

In the data preprocessing stage, all numerical features are first aggregated and normalized to eliminate missing and abnormal values, and downsampling (Zhou, 2020) is used to fill missing values and improve data quality. Based on the three original binary pathological annotation variables, a new multi-classification label ‘Pathologic_typing’ is constructed to characterize the specific pathological classification status of patients and provide a unified prediction target for training downstream multi-classification models.

2.3.2 Feature extraction and screening

In feature extraction and screening, the features of cervical cancer lesions in MRI images of 55 patients with detailed clinical information were extracted by OpenCV (cv2) and the PyRadiomics library. In order to improve the generalization ability of the model and reduce dimension redundancy, a systematic feature selection strategy was used to screen and optimize the input features. First, by evaluating combinations of different numbers of features (ranging from 5 to 100), the performance of each combination was assessed under a cross-validation framework (

Zhong et al., 2010

) to automatically determine the optimal number of features. Three complementary feature selection methods are further introduced:

Based on the model-based selection method, the feature importance score generated by random forest (Qi, 2012) is used to screen high-weight features.

Based on univariate selection, analysis of variance (ANOVA) F-test was used to measure the linear dependence between each feature and the target variable (Chen et al., 2018).

Recursive feature elimination (RFE) (Darst et al., 2018), which gradually converges to the optimal feature subset, was used to iteratively train and remove the least important features.

All features are standardized using the StandardScaler method (Ahmad G. et al., 2022) before model training so that they conform to the standard normal distribution hypothesis with a mean of 0 and a variance of 1, thereby improving the model’s convergence speed and performance stability.

2.3.3 Model training and optimization

This study compared and evaluated the performance of four mainstream classification models in the pathological classification task of cervical cancer, including random forest (RF), extreme gradient boosting (XGBoost, XGB) (Chen et al., 2015), support vector machine (SVM) (Shan, 2016), and logistic regression (LR) (Christodoulou et al., 2019). In order to address the problem of uneven sample distribution, the synthetic minority over-sampling technique (SMOTE) (Chawla et al., 2002) was used to oversample the minority classes in the training set, generating synthetic samples to enhance the model’s ability to recognize underrepresented categories. During the training process, the grid search (GridSearchCV) is used to tune the key hyperparameters of each model (Ahmad G. et al., 2022). Each model presets a specific parameter search space and performs performance evaluation through 5-fold cross-validation to ensure the optimality of the hyperparameter combination in generalization ability (Browne, 2000).

2.3.4 Model evaluation and validation

The performance of the final model is verified in the test set using the following metrics, as defined in Equations 9–12:

True positive (TP) refers to the number of positive samples that are correctly identified as positive, whereas true negative (TN) denotes the number of negative samples correctly predicted as negative. False positive (FP) represents the number of negative samples that are incorrectly classified as positive, and false negative (FN) refers to the number of positive samples that are mistakenly classified as negative. Based on these quantities, the F1-score is adopted as a comprehensive performance indicator.

The F1-score is defined as the harmonic mean of precision and recall, providing a balanced evaluation that simultaneously accounts for both FPs and FNs. Due to its robustness in handling class imbalance, the F1-score offers a more reliable measure of model performance than accuracy alone.

The evaluation metrics, including accuracy, precision, recall, and the F1-score, were calculated (Naidu et al., 2023), and the prediction performance of the model for each category was further analyzed using the confusion matrix. In addition, feature importance analysis was conducted to identify and visualize the features that contributed most to model’s predictions. On this basis, the importance distribution of clinical features and radiomics features was statistically analyzed to reveal their complementary role in the pathological classification task.

2.4 Implementation details

The research was conducted on the Ubuntu 20.04 LTS operating system, using a software environment comprising Python (3.8.10), CUDA (11.7), cuDNN (8.5.0), and the deep learning framework PyTorch (1.13.1) to train and evaluate the model.

3 Result

In this section, we present the experimental results of the proposed framework, which aims to (1) automatically segment cervical cancer lesions using the CRFE model and (2) construct an auxiliary diagnostic model for pathological classification based on radiomics biomarkers. The results are organized into two parts: the performance evaluation of the CRFE segmentation model and the effectiveness of the auxiliary pathological diagnosis model.

3.1 CRFE segmentation model

In terms of segmentation performance indicators, the IoU value of the model is 0.9443, the Dice coefficient is 0.5980, and the F1-score is 0.7085. Among these metrics, the IoU value of 0.9443 is close to the ideal value of 1.0, indicating excellent overall spatial overlap between the predicted and reference lesion regions. Although the Dice coefficient and F1-score are lower, they remain within an acceptable range for medical image segmentation tasks involving complex lesion boundaries. In terms of classification performance, the accuracy rate is 0.8000, the precision rate is 0.9450, the recall rate is 0.8290, and the specificity is 0.6550. In terms of performance indicators, the mean absolute error (MAE) is 70.85, the Hausdorff distance (HD) is 16.46, and the inference time is 450.00 ms. In addition, the model evaluation program generates artificial segmentations of random samples and image comparisons of lesion areas based on the CRFE model’s segmentation. Figures 2A–D present the qualitative segmentation comparisons, while Figure 2E summarizes the quantitative performance metrics.

Segmentation performance of the CRFE model on cervical cancer MRI images. (A) Standard lesion contour of a certain MRI slice; (C) lesion contour obtained by the model after segmentation of the slice; (B) standard lesion area of all the other slices of the patient; (D) lesion contour obtained by the model after segmentation of the slice; (E) performance index of the CRFE model on three test sets of MRI images.

3.2 Auxiliary diagnosis model of pathological classification of cervical cancer3.2.1 Features extracted

Through the integration of multi-source data and the extraction of cervical cancer lesion features, a total of 868 features were obtained. These included clinical features, histogram-based features, morphological features (e.g., Elongation and MajorAxisLength), GLCM (gray-level co-occurrence matrix) features; GLSZM (gray-level size zone matrix) features, GLRLM (gray-level run length matrix) features; GLDM (gray-level dependence matrix) features, NGTDM (neighborhood gray-tone difference matrix) features; and first-order statistical features (Table 3). After feature selection, 30 features with the highest predictive power were retained from the original feature set (Tables 4–6). Analysis of feature importance and feature correlation revealed that these 30 selected features included a histogram-based feature, namely, the median, which represents the median pixel intensity and ranked eighth in importance among all the features. The feature correlation heatmap more clearly illustrates the relationships among the features, indicating that the 30 selected features—including histogram features derived from MRI images of the lesion regions—are closely associated with the auxiliary diagnosis of cervical cancer (Figures 3, 4). The cutoff at the top 30 features was determined based on the distribution of normalized feature importance scores. As shown in Figure 3, the importance values exhibited a clear inflection point after the 30th feature, with scores decreasing sharply below 0.01. Including additional features beyond this point did not lead to further improvement in classification performance during cross-validation, while increasing the risk of feature redundancy and overfitting. The feature correlation heatmap in Figure 4 shows distinct groups of radiomics features with strong internal correlations. Features derived from the same texture matrices and similar filtering strategies tend to cluster together, indicating shared information content. In particular, GLSZM- and GLRLM-based features form tightly correlated clusters, indicating that they capture complementary aspects of lesion heterogeneity, while first-order and histogram features show weaker correlations with texture-based features, highlighting their independent contribution to the classification model.

Feature categoryDescriptionNumber of featuresClinical featuresPatient demographics and pathological indicators, including age, HPV status, FIGO stage, lymph-node metastasis, lymphovascular invasion, and treatment information8Morphological featuresShape- and geometry-related features extracted from lesion regions, including volume, surface area, elongation, major axis length, and sphericity14First-order statistical featuresIntensity-based features describing voxel intensity distributions, such as the mean, median, variance, skewness, kurtosis, and root mean squared values162Texture features (GLCM)Gray-level co-occurrence matrix features characterizing spatial gray-level dependencies, including contrast, correlation, entropy, and homogeneity216Texture features (GLRLM)Gray-level run length matrix features describing the run-length distributions of gray levels144Texture features (GLSZM)Gray-level size zone matrix features quantifying homogeneous gray-level zones144Texture features (GLDM)Gray-level dependence matrix features measuring gray-level dependence patterns126Texture features (NGTDM)Neighborhood gray-tone difference matrix features reflecting local gray-level differences45Histogram featuresHistogram-based features describing gray-level distribution trends9Total​868

Summary of the extracted features used for the auxiliary diagnosis model.

Bold values indicate the total number of extracted features.

Feature nameDescriptionGLRLM long run low gray-level emphasis (LoG, σ = 2.0 mm, 3D)Emphasis on long continuous segments of low gray-levels, extracted from 3D images after Laplacian of Gaussian filtering with σ = 2.0 mmGLSZM large area low gray-level emphasis (LoG, σ = 2.0mm, 3D)Emphasis on large-sized regions with low gray-levels, extracted from 3D images after LoG filtering (σ = 2.0 mm)NGTDM strength (LoG, σ = 1.0 mm, 3D)Average gray-tone difference in local neighborhoods, reflecting the intensity of local gray-level variations (σ = 1.0 mm)GLSZM large area high gray-level emphasis (LoG, σ = 2.0 mm, 3D)Emphasis on large-sized regions with high gray-levels, extracted after LoG filtering (σ = 2.0 mm)GLDM dependence variance (LoG, σ = 1.0 mm, 3D)Variance of gray-level dependencies (correlations at fixed distance) with σ = 1.0 mmGLRLM short-run low gray-level emphasis (Wavelet-L)Emphasis on short continuous runs of low gray-levels from low-frequency wavelet componentsGLCM difference variance (Wavelet-L)Variance of gray-level differences, reflecting dispersion of gray-level changes in low-frequency waveletsMedian of histogramMedian value of gray-level histogram, representing central tendency of distribution