Non-invasive endometriosis staging prediction using integrated radiomics and spatiotemporal transformer model based on dynamic contrast-enhanced MRI

Precise staging of endometriosis remains a clinical challenge, as current diagnosis depends almost entirely on laparoscopic visualization—an invasive procedure marked by considerable inter- observer disagreement and diagnostic delays. Existing non-invasive approaches, whether based on conventional machine learning with handcrafted features or end-to-end deep learning architectures, have shown limited success in capturing the complex spatiotemporal characteristics of endometriotic lesions on dynamic imaging. Traditional radiomics methods struggle to model temporal enhancement patterns, while pure deep learning approaches often overlook domain-specific textural signatures that correlate with disease severity. To address these limitations, we developed a computational framework that combines the interpretability of ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS engineered radiomics features with the representational power of spatiotemporal Transformers, specifically designed for dynamic contrast-enhanced magnetic resonance imaging sequences. Our dual- pathway architecture employs cross-modal attention mechanisms that allow bidirectional refinement between texture-based descriptors and learned spatiotemporal representations, alongside adaptive weighting that adjusts feature contributions based on individual case characteristics. When evaluated on a multi-institutional cohort of 486 surgically confirmed cases, the model achieved 87.8% accuracy with a macro-averaged F1-score of 0.854 on the test set, substantially outperforming both traditional machine learning baselines and single- modality deep networks. Independent validation on 127 prospective patients from an external center demonstrated consistent performance with 85.0% accuracy, suggesting reasonable generalization across different clinical settings. Attention weight visualizations revealed that the model appropriately focuses on anatomical structures documented in surgical staging criteria, including ovarian endometriomas, deep infiltrating sites, and peritoneal implants. By providing interpretable staging predictions with quantified confidence levels, this framework may support preoperative risk stratification and assist clinicians in identifying patients who would benefit most from surgical intervention, though its ultimate impact on treatment decision-making requires prospective evaluation.

Endometriosis staging; Radiomics; Spatiotemporal Transformer; Deep learning; Dynamic contrast-enhanced MRI; Non-invasive prediction I. Introduction Endometriosis affects roughly 10% of women during their reproductive years, characterized by ectopic endometrial tissue growth that triggers chronic pelvic pain, dysmenorrhea, dyspareunia, and infertility—symptoms that profoundly diminish quality of life [1, 2]. Staging this condition has long posed a clinical dilemma. The revised American Society for Reproductive Medicine (rASRM) classification, introduced in 1996, remains the most widely adopted system for surgical documentation and standardized communication among clinicians [11]. Yet this scoring framework carries well- ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS recognized limitations that affect its clinical utility. Most notably, the rASRM system emphasizes superficial peritoneal disease and ovarian endometriomas while inadequately representing deep infiltrating endometriosis (DIE), which often drives the most severe symptoms and surgical complexity. Studies have repeatedly shown weak correlations between rASRM stage and key clinical outcomes such as pain severity, fertility potential, and surgical difficulty [7]. In response to these shortcomings, complementary classification systems have emerged, including the Enzian classification for deep disease and the Endometriosis Fertility Index (EFI) for predicting reproductive outcomes. Despite these alternatives, rASRM continues to serve as the primary staging reference in most clinical and research contexts, creating a persistent need for tools that can work within this established framework while acknowledging its constraints. Current diagnostic practice still relies heavily on laparoscopic visualization combined with histopathological confirmation—an invasive approach that carries surgical risks, generates substantial costs, and demonstrates considerable inter-observer variation in lesion interpretation and stage assignment [3, 13]. Moreover, laparoscopy cannot reliably detect microscopic peritoneal implants or predict functional outcomes, underscoring the clinical value of developing accurate preoperative staging tools based on non-invasive imaging. Radiomics offers a quantitative approach to disease characterization by extracting high-dimensional features that describe texture, morphology, and intensity distributions within medical images [4, 14]. This methodology has shown promise across gynecological imaging, including applications in ovarian cancer risk assessment, fibroid characterization, and adenomyosis detection [4]. Radiomic descriptors can reveal subtle spatial heterogeneity patterns that escape visual inspection, potentially capturing biological signatures related to inflammation, fibrosis, and vascular remodeling. However, conventional radiomics pipelines depend entirely on predefined feature engineering—a strategy that may prove insufficient for modeling the dynamic vascular changes visible in contrast-enhanced sequences. Recent work has attempted to apply radiomics to endometriosis, yet these studies typically analyze static imaging snapshots and report modest diagnostic accuracy, particularly for advanced disease stages where lesion heterogeneity is most pronounced. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Deep learning has reshaped medical imaging analysis, with Transformer architectures emerging as particularly powerful tools for modeling complex dependencies in visual data [5]. Unlike convolutional neural networks that process images through localized receptive fields, Transformers employ self-attention mechanisms that establish direct connections between distant image regions, enabling holistic integration of contextual information [5, 6]. Recent applications have extended Transformers to dynamic medical imaging sequences, where spatiotemporal attention modules can jointly model anatomical structures and their evolution over time [22, 23]. Several studies published between 2022 and 2024 have demonstrated that hybrid architectures combining Transformers with domain-specific feature extraction yield superior performance compared to purely data-driven approaches, particularly in scenarios with limited training data—a common constraint in medical imaging research [54, 55, 56, 57, 58]. For endometriosis specifically, dynamic contrast-enhanced MRI captures temporal enhancement kinetics that reflect lesion vascularity, inflammation status, and fibrotic tissue composition, yet existing deep learning studies have predominantly analyzed static sequences and failed to exploit this temporal dimension [51, 52, 53]. This represents a substantial missed opportunity, as enhancement patterns evolve characteristically across disease stages and may provide discriminative signals beyond what static morphological features can capture. Several interconnected challenges have limited progress in computational endometriosis staging. First, endometriotic lesions exhibit remarkable phenotypic heterogeneity, with individual patients presenting variable combinations of superficial implants, ovarian cysts, and deep infiltrating nodules that differ in signal characteristics and morphological appearance [7, 10]. This biological diversity complicates feature-based classification, as no single imaging signature consistently predicts disease stage. Second, most prior computational approaches have treated staging as a static image classification problem, applying conventional machine learning algorithms to handcrafted features extracted from single imaging time points. While these radiomics-based methods achieve reasonable performance for early-stage disease, they struggle with advanced cases where lesion complexity increases and sample sizes decrease—a fundamental limitation of engineered features that lack the capacity ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS to learn hierarchical representations from data. Third, recent deep learning studies have demonstrated that convolutional networks can automatically extract discriminative features, yet these models often require large training cohorts and provide limited interpretability, making clinical adoption difficult. Perhaps most critically, no prior work has systematically integrated the complementary strengths of radiomics and deep learning while simultaneously modeling the temporal enhancement dynamics captured in dynamic contrast- enhanced sequences. This gap motivated our investigation into hybrid architectures that could leverage both domain knowledge encoded in handcrafted features and data-driven representations learned through attention mechanisms designed specifically for spatiotemporal medical imaging. We developed a computational framework designed to address these

through synergistic integration of radiomics and spatiotemporal Transformer architectures. Rather than relying exclusively on either handcrafted or learned features, our approach combines both paradigms through a cross-modal attention mechanism that allows bidirectional refinement—texture-based radiomics descriptors inform the Transformer’s attention weights, while learned spatiotemporal representations help reweight radiomic features based on their relevance to each case. This fusion strategy aims to capture both the fine-grained tissue characteristics that radiologists recognize visually and the complex temporal patterns that emerge during contrast enhancement. Our work introduces several technical contributions: first, a hybrid feature extraction pathway that processes static radiomics and dynamic imaging sequences in parallel before fusing them through learned attention; second, a divided space-time attention module that efficiently models both intra-frame anatomical relationships and inter-frame temporal evolution; third, adaptive weighting that adjusts the contribution of each feature modality based on prediction confidence; and fourth, a class-balanced training strategy using focal loss to address the inherent data imbalance across staging categories. If validated through prospective clinical studies, such a framework could potentially support preoperative risk stratification by identifying patients most likely to have advanced disease, thereby informing decisions about surgical planning, multidisciplinary consultation, and fertility preservation— though we emphasize that determining its actual impact on clinical ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS workflows and patient outcomes requires rigorous evaluation beyond the scope of this retrospective analysis [8]. Our primary objective in this work was to establish whether integrating complementary feature representations with spatiotemporal modeling could improve staging accuracy over existing computational approaches, while providing sufficient interpretability to support clinical trust and facilitate future translation studies. II. Related Theory and Technical Foundation 2.1 Clinical Features and Staging System of Endometriosis Multiple mechanisms contribute to endometriosis pathogenesis, though retrograde menstruation—where viable endometrial cells travel through fallopian tubes and implant on peritoneal surfaces— remains the most widely invoked explanation [9]. Alternative hypotheses involving coelomic metaplasia, lymphatic dissemination, and stem cell transformation continue to be investigated. Clinical presentations span a remarkably wide spectrum, from asymptomatic findings discovered incidentally during unrelated procedures to severe pain syndromes that profoundly disrupt daily functioning and reproductive capacity. Deep infiltrating disease, which penetrates more than 5mm beneath the peritoneal surface, commonly affects the rectovaginal septum, uterosacral ligaments, and bowel serosa, manifesting with intractable dysmenorrhea, painful defecation, and infertility that may persist despite medical management [10]. The rASRM classification divides endometriosis into four stages through a weighted point system: Stage I (minimal: 1-5 points) and Stage II (mild: 6-15 points) typically feature superficial peritoneal implants with minimal adhesions; Stage III (moderate: 16-40 points) includes multiple deep implants, ovarian endometriomas, and filmy adhesions; Stage IV (severe: >40 points) encompasses extensive deep disease with dense adhesions and large bilateral ovarian cysts [11]. Points accumulate based on lesion size, depth, location, and adhesion severity, offering standardized surgical documentation. However, this system has important constraints that affect its clinical interpretation. The rASRM score correlates poorly with symptom severity—patients with minimal Stage I disease may experience debilitating pain, while some with Stage IV disease remain relatively asymptomatic. The classification emphasizes visible peritoneal and ovarian pathology but ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS inadequately captures deep infiltrating lesions, which often determine surgical complexity and symptom burden more than rASRM stage alone. Furthermore, the system does not incorporate factors like lesion proximity to critical structures (ureters, bowel, major vessels) or predict fertility outcomes reliably. These limitations led to development of complementary systems: the Enzian classification specifically addresses deep retroperitoneal disease, while the Endometriosis Fertility Index attempts to predict post-surgical pregnancy rates by incorporating age, fertility history, and surgical findings [7]. Despite these alternatives, rASRM persists as the dominant staging reference in clinical practice and research, making it the pragmatic target for predictive modeling even while acknowledging that stage alone provides incomplete prognostic information. Current imaging modalities for endometriosis evaluation include transvaginal ultrasound and magnetic resonance imaging. Transvaginal ultrasound offers real-time assessment with high spatial resolution for detecting ovarian endometriomas, while MRI provides superior soft tissue contrast and multiplanar imaging capability for evaluating deep infiltrating lesions [12]. T1-weighted fat-suppressed sequences demonstrate characteristic high signal intensity in endometriomas due to hemorrhagic content, and T2-weighted images reveal fibrous tissue and implant architecture. Traditional staging diagnosis relies on direct laparoscopic visualization combined with histopathological confirmation, which remains the gold standard despite inherent limitations. This invasive approach carries surgical risks, incurs substantial healthcare costs, and exhibits significant inter-observer variability in lesion assessment and stage assignment [13]. Moreover, laparoscopy cannot reliably detect microscopic peritoneal disease or predict functional outcomes, highlighting the necessity for developing accurate non-invasive predictive models based on preoperative imaging data. Our framework specifically employs dynamic contrast-enhanced MRI (DCE-MRI), which captures temporal enhancement patterns by acquiring sequential images after intravenous gadolinium administration. At our participating institutions, DCE sequences have been incorporated into standard endometriosis MRI protocols for patients with suspected deep infiltrating disease, based on the rationale that enhancement kinetics may reflect lesion vascularity and ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS inflammatory activity. However, we acknowledge that DCE-MRI is not universally adopted across all centers—many institutions rely primarily on T2-weighted and T1-weighted fat-suppressed sequences without dynamic contrast. This creates potential barriers to widespread implementation of DCE-dependent models, particularly in resource-limited settings where extended scan times, additional contrast administration, and associated costs may not be feasible. The trade-offs are non-trivial: DCE imaging adds approximately 8-12 minutes to total examination time, requires intravenous access and contrast injection with attendant (albeit low) risks of allergic reactions and nephrogenic concerns, and generates substantially larger datasets that demand more storage and processing capacity. Future work should explore whether comparable staging accuracy might be achieved using conventional non-contrast sequences alone, potentially through architectural modifications or alternative temporal modeling strategies applied to multi-planar acquisitions, thereby broadening the model’s applicability to diverse clinical environments. 2.2 Radiomics Theory and Feature Extraction Methods Radiomics represents a high-throughput approach that converts medical images into mineable quantitative data by extracting a large number of computational features describing tissue texture, shape, and intensity patterns [14]. The standard radiomics workflow comprises image acquisition and reconstruction, region of interest segmentation, feature extraction, feature selection, and predictive model construction. This pipeline enables objective quantification of phenotypic characteristics that correlate with underlying pathophysiological processes, thereby facilitating precision diagnosis and prognosis prediction. Radiomic features fall into several conceptually distinct categories, each capturing different aspects of tissue appearance. First-order statistics describe the overall intensity distribution—mean, variance, skewness, and kurtosis summarize how bright or dark a lesion appears and whether its intensity values are evenly distributed or skewed [15]. Shape features quantify three-dimensional geometry: volume, surface area, sphericity, and compactness describe whether lesions are large or small, smooth or irregular, round or elongated. Second-order texture features probe spatial patterns by analyzing how often certain intensity combinations appear at specific distances ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS and directions—these are computed through matrices like GLCM, GLRLM, and GLSZM that encode local neighborhood relationships. Higher-order features from wavelet decomposition and Laplacian-of- Gaussian filtering capture information at multiple scales, revealing both coarse structural patterns and fine edge details that may correlate with tissue composition and architectural disorganization [16]. Mathematically, these features transform raw voxel intensities into numerical descriptors through well-defined operations—for instance, GLCM-based contrast measures local intensity differences by weighting co-occurrence probabilities, while sphericity compares a structure’s surface-to-volume ratio against that of a perfect sphere. The full mathematical formulations of these features are documented in standardized radiomics references and implementations like PyRadiomics [15]. Feature selection and dimensionality reduction constitute critical steps to mitigate overfitting and enhance model interpretability. Common approaches include univariate statistical tests, correlation analysis, recursive feature elimination, and regularization techniques such as Least Absolute Shrinkage and Selection Operator (LASSO) [17]. Principal component analysis and independent component analysis transform original features into uncorrelated components, reducing computational complexity while preserving variance. The diagnostic application of radiomics leverages machine learning algorithms to establish mappings between extracted features and clinical outcomes. Radiomic signatures can capture tumor heterogeneity, microenvironmental characteristics, and treatment response patterns invisible to conventional imaging interpretation [18]. By integrating handcrafted radiomic features with clinical variables and molecular biomarkers, comprehensive predictive models achieve improved diagnostic accuracy and risk stratification in oncological and non-oncological diseases. 2.3 Spatiotemporal Transformer Model Architecture The Transformer architecture revolutionized sequence modeling by replacing recurrent connections with self-attention mechanisms, enabling parallel computation and capturing long-range dependencies effectively [19]. The core innovation lies in the attention mechanism that computes weighted representations of input sequences, where ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS each element attends to all other elements regardless of their positional distance. This property makes Transformers particularly suitable for modeling complex spatiotemporal patterns in medical imaging sequences. Self-attention computes relationships between all positions in a sequence by transforming inputs into query, key, and value representations, then calculating attention weights based on query- key similarity and using these weights to aggregate values [19, 20]. This differs fundamentally from convolutional operations, which process only local neighborhoods—attention can establish direct connections between distant image regions in a single step, enabling holistic integration of context. The mechanism employs learned projection matrices to create queries, keys, and values, computes similarity scores between queries and all keys, normalizes these scores through a softmax function to produce attention weights, then uses these weights to create a weighted combination of values. A scaling factor prevents numerical instability when feature dimensions are large. Multi-head attention extends this concept by learning multiple parallel attention functions, each potentially capturing different types of relationships—one head might focus on anatomical boundaries, another on intensity patterns, a third on spatial proximity [21]. Outputs from all heads are concatenated and projected back to the original dimensionality, allowing the model to jointly attend to information from multiple representational subspaces. Positional encoding injects sequential order information into the model, as self-attention operations are inherently permutation- invariant. Sinusoidal position encoding employs fixed functions with varying frequencies, while learnable positional embeddings optimize position representations during training. For medical imaging applications, three-dimensional spatial coordinates and temporal frame indices require careful encoding to preserve anatomical and temporal relationships. For medical imaging sequences that evolve over time, spatiotemporal Transformers must model both how anatomical structures relate within each frame and how these structures change across frames [22]. A naive approach would compute attention across all spatiotemporal positions jointly, but this quickly becomes computationally prohibitive as sequence length increases—the ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS complexity scales quadratically with the total number of space-time positions. Divided space-time attention offers a practical solution by factorizing the problem: first computing spatial attention within each temporal frame to model anatomical relationships, then computing temporal attention across frames for each spatial location to capture dynamics [23]. This factorization reduces complexity substantially while still allowing information to flow both spatially and temporally through the network. For dynamic contrast-enhanced MRI, this architecture proves particularly suitable—spatial attention can identify lesion boundaries and their relationship to surrounding anatomy, while temporal attention captures enhancement kinetics that reflect vascular permeability and tissue perfusion, patterns that evolve characteristically across disease stages. III. Integrated Radiomics and Spatiotemporal Transformer Model for Staging Prediction 3.1 Overall Model Framework Design The proposed model adopts a dual-pathway architecture that synergistically integrates handcrafted radiomics features with automatically learned spatiotemporal representations to achieve accurate endometriosis staging prediction. Figure 1 illustrates the complete workflow, encompassing image preprocessing, parallel feature extraction through radiomics and Transformer pathways, feature fusion, and staging classification. This hybrid design leverages the complementary strengths of domain-specific engineered features and data-driven deep learning representations [24]. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Figure 1. Overall architecture of the integrated radiomics and spatiotemporal Transformer model for endometriosis staging prediction. The image preprocessing pipeline standardizes multi-sequence MRI data to ensure consistency across different scanning protocols and institutions. Raw DICOM images undergo intensity normalization using N4 bias field correction to eliminate signal inhomogeneity, followed by histogram matching to align intensity distributions across ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS subjects [25]. Spatial resampling to isotropic voxel spacing of 1.0 × 1.0 × 1.0 mm³ maintains anatomical proportions and facilitates subsequent analysis. Region of interest delineation encompasses lesion areas, ovarian structures, and deep infiltrating sites, performed either manually by experienced radiologists or semi-automatically through segmentation networks with expert refinement. The radiomics feature extraction module employs PyRadiomics toolkit to compute comprehensive quantitative descriptors from preprocessed images. First-order statistical features, shape-based morphological features, and second-order texture features derived from GLCM, GLRLM, GLSZM, and neighboring gray-tone difference matrices are systematically extracted. Wavelet decomposition generates multi-scale representations across eight frequency subbands, expanding the feature space to capture fine-grained tissue heterogeneity. To prevent data leakage and ensure valid performance estimation, all feature selection and normalization steps were performed exclusively within the training set using five-fold cross- validation—specifically, we first removed features with near-zero variance (standard deviation 0.95) to reduce redundancy, then applied LASSO regression with cross-validated alpha tuning to select the most discriminative subset [26]. The normalization parameters (mean and standard deviation for z-score standardization) were computed solely from training data, then applied to validation and test sets without any reverse information flow. Similarly, the LASSO regularization path and optimal lambda parameter were determined entirely within training folds, with the final selected feature subset (25 features from an initial pool of 1,316) fixed before any evaluation on held-out data. The validation set was used only for monitoring training progress and early stopping decisions, while the test set and external prospective cohort remained completely unseen until final model evaluation. This strict separation ensures that reported performance metrics reflect genuine generalization capability rather than overfitting artifacts from inadvertent information leakage. The spatiotemporal Transformer encoding module processes dynamic contrast-enhanced sequences to model temporal evolution patterns of lesion enhancement. Input image sequences are partitioned into non- overlapping three-dimensional patches and linearly embedded into ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS token representations. Learnable positional embeddings encode spatial coordinates and temporal frame indices, preserving anatomical topology and sequential order. Stacked Transformer encoder layers with divided space-time attention progressively extract hierarchical spatiotemporal features, where spatial self-attention captures intra- frame anatomical relationships and temporal self-attention models inter-frame dynamics. Layer normalization and residual connections stabilize training and facilitate gradient flow through deep architectures. Table 1 summarizes the detailed parameter configuration for each module component. The model employs 6 Transformer encoder layers with 8 attention heads, achieving a balance between representational capacity and computational efficiency. The patch size of 16 × 16 × 4 (H × W × T) provides adequate spatial resolution while maintaining manageable sequence length for attention computation. Table 1. Parameter configuration of model components Module Component Parameter Name Value Dimension Description Image Preprocessing Voxel Spacing 1.0 × 1.0 × 1.0 mm³ 3D Isotropic resolution Radiomics Extraction Feature Number 1,316 1D Total extracted features Radiomics Selection Selected Features 25 1D Post-LASSO selection Transformer Patch Patch Size 16 × 16 × 4 H × W × T Spatial- temporal patch Transformer Encoder Number of Layers 6 Scalar Encoder depth Transformer Encoder Attention Heads 8 Scalar Multi-head attention Transformer Encoder Embedding Dimension 512 1D Feature dimension Feature Fusion Fusion Dimension 256 1D Combined feature space The feature fusion strategy employs concatenation followed by fully connected projection to integrate radiomics and Transformer-derived representations into a unified feature space. Cross-attention mechanisms optionally enable bidirectional interaction between the two feature modalities, allowing Transformer features to refine radiomics representations and vice versa. The staging prediction output layer consists of multiple fully connected layers with dropout ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS regularization, culminating in a softmax classifier that generates probability distributions over four rASRM staging categories [27]. Class-balanced focal loss function addresses the inherent data imbalance across different disease stages, emphasizing learning from hard-to-classify samples. The complete model architecture comprises approximately 47.3 million trainable parameters distributed across its major components: the radiomics feature processing pathway contains 0.8 million parameters (multilayer projection networks and normalization layers), the spatiotemporal Transformer encoder accounts for 41.2 million parameters (six encoder layers each with multi-head attention, feedforward networks, and layer normalization), and the cross-modal fusion module with adaptive weighting adds 5.3 million parameters (cross-attention mechanisms and gating networks). The full model occupies 181 MB of storage in half-precision (FP16) format and 362 MB in single-precision (FP32) format when including both architecture definition and trained weights. While these parameter counts are substantial, they remain manageable compared to large- scale vision models—for context, Vision Transformer Base contains 86 million parameters and ResNet-152 contains 60 million parameters. Memory requirements during inference are approximately 4.2 GB for batch size 1 on GPU, enabling deployment on consumer-grade hardware such as NVIDIA RTX 2080 or higher. Training the full model required approximately 18 hours on a single NVIDIA RTX 3090 GPU, with convergence typically achieved within 120-150 epochs depending on random initialization. These computational characteristics suggest that while the model demands more resources than traditional machine learning approaches, it remains practical for deployment in radiology departments with modern workstation hardware. 3.2 Spatiotemporal Feature Fusion Mechanism The spatiotemporal feature fusion mechanism constitutes a critical component that enables effective integration of heterogeneous feature representations derived from radiomics and Transformer pathways. Figure 2 depicts the architecture of the proposed fusion module, which employs cross-modal attention to establish semantic correspondences between handcrafted and learned features while preserving their complementary characteristics. This design addresses the fundamental challenge of combining features from ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS disparate extraction paradigms with different statistical properties and representational capacities [28]. Figure 2. Architecture of the spatiotemporal attention fusion module showing cross-modal feature alignment and adaptive weight learning. The spatiotemporal representation of radiomics features transforms static handcrafted descriptors into dynamic embeddings compatible with Transformer-derived representations. Given a radiomics feature vector fr ∈ Rdr extracted from each temporal frame, temporal pooling operations aggregate features across dynamic sequences to construct frame-wise representations. A learnable temporal projection network maps these aggregated features into a sequence Fr = {f1r,f2r,...,fTr} ∈ RT×dr, where T denotes the number of temporal frames. This transformation enables radiomics features to participate in temporal attention mechanisms alongside Transformer features, facilitating coherent spatiotemporal modeling [29]. Cross-modal feature alignment establishes bidirectional interactions between radiomics representations Fr and Transformer-encoded ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS features Ft ∈ RT×dt through dual cross-attention operations. The radiomics-to-Transformer attention computes: Fr→t = softmax( FrWQ(FtWK)T dk )FtWV (8) where WQ,WK,WV represent learnable projection matrices. This operation allows radiomics features to query relevant spatiotemporal patterns from Transformer representations, capturing dynamic enhancement characteristics informed by texture descriptors. Conversely, Transformer-to-radiomics attention Ft→r enables deep features to refine handcrafted representations by incorporating learned semantic context. The bidirectional alignment mechanism ensures mutual enhancement between the two modalities while maintaining their distinct representational advantages [30]. Adaptive weight learning dynamically adjusts the contribution of each feature modality based on input characteristics through a gating mechanism. The fusion weights αr and αt for radiomics and Transformer features respectively are computed via a small neural network: αr,αt = softmax(MLP([Fpoolr ;Fpoolt ])) (9) where Fpoolr and Fpoolt denote global average pooled representations and [·;·] indicates concatenation. This adaptive weighting accommodates heterogeneous data quality and varying discriminative power across different cases, assigning higher weights to more reliable feature modalities for each individual prediction. The mathematical formulation of the complete fusion process combines aligned features and adaptive weights to generate the unified representation: Ffused= αr·(Fr + Ft→r) + αt·(Ft + Fr→t) (10) ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS This formulation incorporates both original and cross-attended features from each modality, weighted by learned importance coefficients. Residual connections preserve gradient flow and prevent feature degradation during fusion. The fused representation Ffused encapsulates comprehensive spatiotemporal characteristics spanning handcrafted texture patterns and automatically learned semantic abstractions, providing a robust foundation for subsequent staging classification [31]. The fusion mechanism exhibits computational efficiency through parallel processing of cross-attention operations and lightweight gating networks. Dropout regularization applied to attention weights prevents overfitting to spurious correlations between modalities, enhancing model generalization to unseen data distributions. Layer normalization stabilizes feature magnitudes across different modalities, facilitating stable gradient-based optimization during training. 3.3 Model Training and Optimization Strategy The loss function design addresses the inherent class imbalance in endometriosis staging datasets, where advanced stages typically comprise fewer cases than early stages. Focal loss extends conventional cross-entropy by down-weighting well-classified examples and focusing on hard-to-classify samples [32]. The mathematical formulation is: Lfocal= - C ∑ c=1 αc(1 - pc)γyclog(pc) (11) where C denotes the number of staging classes, yc represents the ground truth label, pc is the predicted probability for class c, αc adjusts class weights inversely proportional to class frequency, and γ controls the focusing strength. The modulating factor (1 - pc)γ reduces loss contribution from easily classified examples, directing optimization toward challenging cases. Additional L2 regularization on model parameters prevents excessive weight magnitudes and improves generalization. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS The optimization algorithm employs AdamW optimizer, which decouples weight decay from gradient-based updates to achieve superior regularization compared to standard Adam [33]. AdamW maintains adaptive learning rates for individual parameters through first and second moment estimates while applying weight decay directly to parameters. This decoupling prevents weight decay from being scaled by adaptive learning rates, leading to more effective regularization and improved convergence in deep architectures. Learning rate scheduling adopts a cosine annealing strategy with linear warmup to stabilize early-stage training and enable fine- grained optimization in later epochs. The learning rate ηt at epoch t follows: ηt = ηmin+ 1 2 (ηmax- ηmin)(1 + cos( t - twarmup Tmax- twarmup π)) (12) where ηmax and ηmin denote maximum and minimum learning rates, twarmup represents warmup epochs, and Tmax indicates total training epochs. Linear warmup gradually increases learning rate from near- zero to ηmax during initial epochs, preventing destabilization from large gradients when parameters are randomly initialized [34]. Data augmentation techniques enhance model robustness and mitigate overfitting by expanding training sample diversity. Spatial transformations include random rotation within ±15 degrees, random scaling by 0.9-1.1 factors, and elastic deformation with controlled displacement fields. Intensity augmentation applies random contrast adjustment, Gaussian noise injection with 0.01 standard deviation, and gamma correction to simulate imaging variability across scanners and protocols. Temporal augmentation randomly samples subsequences from dynamic series and applies temporal jittering to improve invariance to acquisition timing variations. Table 2 lists the comprehensive hyperparameter configuration employed during model training. The batch size of 16 balances GPU memory constraints with gradient estimate stability, while 200 training epochs ensure sufficient convergence with early stopping patience of 30 epochs to prevent overfitting. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Table 2. Training hyperparameter settings Hyperparameter Value Description Batch Size 16 Training batch size Initial Learning Rate 1e-4 Starting learning rate Weight Decay 1e-5 L2 regularization coefficient Dropout Rate 0.3 Dropout probability Training Epochs 200 Maximum training epochs Early Stopping Patience 30 Epochs without improvement Overfitting prevention employs multiple complementary strategies beyond data augmentation. Dropout with 0.3 probability randomly deactivates neurons during training, preventing co-adaptation and encouraging robust distributed representations. Early stopping monitors validation loss and terminates training when no improvement occurs for 30 consecutive epochs, preserving the best- performing model checkpoint. Gradient clipping constrains gradient norms to prevent exploding gradients in deep networks. Model evaluation adopts a comprehensive metric system encompassing classification accuracy, class-specific precision and recall, macro-averaged F1-score, and area under the receiver operating characteristic curve (AUC) for each staging category [35]. Cohen's kappa coefficient quantifies agreement between predictions and ground truth while accounting for chance agreement. These metrics provide multifaceted assessment of model performance across different staging classes, particularly crucial given class imbalance considerations. IV. Experimental Design and Results Analysis 4.1 Dataset Construction and Experimental Setup The dataset comprises dynamic contrast-enhanced MRI examinations collected from three tertiary hospitals between January 2019 and December 2023, following institutional review board approval and patient informed consent procedures. Imaging was performed on 1.5T and 3.0T MRI systems from multiple vendors (Siemens, GE Healthcare, Philips), with protocols designed to balance image quality and examination time constraints typical of clinical pelvic imaging [36]. All examinations included T1-weighted, T2-weighted, and DCE sequences acquired in axial, sagittal, and coronal planes covering the entire pelvis from the pelvic brim to the perineum. Slice thickness ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS ranged from 3 to 5 mm with no interslice gap, matrix sizes varied from 256×256 to 512×512, and field of view was adjusted to patient body habitus (typically 280-400 mm). DCE sequences employed T1- weighted gradient echo acquisitions with temporal resolution of 30-60 seconds, acquiring 6-8 post-contrast phases following intravenous administration of gadolinium-based contrast agents (gadobutrol or gadoterate meglumine) at standard doses of 0.1 mmol/kg body weight with injection rates of 2-3 mL/second. Detailed scanner-specific parameters for each center are provided in Supplementary Table 1. To address heterogeneity in temporal resolution across sites, we implemented a standardized preprocessing pipeline: DCE sequences with higher temporal resolution (acquisitions every 30 seconds) were downsampled to match those with lower resolution (every 60 seconds) through averaging of consecutive time points, while all sequences were temporally aligned to a reference timing scheme starting at contrast injection and extending through delayed phases. Intensity normalization applied N4 bias field correction followed by z-score standardization within each examination to mitigate scanner- dependent signal variations, and spatial resampling to isotropic 1.0×1.0×1.0 mm³ voxels ensured consistent anatomical scaling across all cases regardless of original acquisition parameters. Sample inclusion criteria required confirmed endometriosis diagnosis through laparoscopic visualization and histopathological examination within 30 days of MRI acquisition, comprehensive surgical staging documentation according to rASRM classification, complete dynamic contrast-enhanced imaging sequences without significant motion artifacts, and availability of clinical records including symptom severity and laboratory results. Exclusion criteria eliminated cases with prior surgical treatment for endometriosis, concurrent malignancies, incomplete imaging protocols, severe motion artifacts compromising image quality, and missing or ambiguous surgical staging information [37]. Table 3 summarizes the statistical characteristics of the final dataset. The cohort encompasses 486 patients with confirmed endometriosis staging, demonstrating the expected class imbalance with decreasing sample sizes for advanced stages. The mean age of 34.2 ± 6.8 years reflects the typical demographic of endometriosis patients during reproductive years. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Table 3. Baseline characteristics and clinical features of the study cohort Characteristic Training Set (n=291) Validation Set (n=97) Test Set (n=98) Total (n=486) rASRM Stage Distribution Stage I (Minimal) 95 (32.6%) 32 (33.0%) 33 (33.7%) 160 (32.9%) Stage II (Mild) 89 (30.6%) 30 (30.9%) 30 (30.6%) 149 (30.7%) Stage III (Moderate) 68 (23.4%) 22 (22.7%) 23 (23.5%) 113 (23.3%) Stage IV (Severe) 39 (13.4%) 13 (13.4%) 12 (12.2%) 64 (13.2%) Demographics Age (years), mean ± SD 34.1 ± 6.9 34.5 ± 6.5 34.3 ± 7.1 34.2 ± 6.8 BMI (kg/m²), mean ± SD 22.8 ± 3.4 23.1 ± 3.2 22.6 ± 3.6 22.8 ± 3.4 Clinical Symptoms, n (%) Dysmenorrhea (moderate-severe) 234 (80.4%) 79 (81.4%) 77 (78.6%) 390 (80.2%) Chronic pelvic pain 198 (68.0%) 64 (66.0%) 69 (70.4%) 331 (68.1%) Dyspareunia 147 (50.5%) 51 (52.6%) 48 (49.0%) 246 (50.6%) Infertility (primary or secondary) 189 (64.9%) 61 (62.9%) 65 (66.3%) 315 (64.8%) Comorbidities, n (%) Adenomyosis 87 (29.9%) 27 (27.8%) 31 145 ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Characteristic Training Set (n=291) Validation Set (n=97) Test Set (n=98) Total (n=486) (31.6%) (29.8%) Uterine fibroids 72 (24.7%) 25 (25.8%) 23 (23.5%) 120 (24.7%) Previous pelvic surgery 58 (19.9%) 18 (18.6%) 21 (21.4%) 97 (20.0%) Previous endometriosis surgery 0 (0%) 0 (0%) 0 (0%) 0 (0%) MRI Scanner Field Strength, n (%) 1.5 Tesla 162 (55.7%) 52 (53.6%) 56 (57.1%) 270 (55.6%) 3.0 Tesla 129 (44.3%) 45 (46.4%) 42 (42.9%) 216 (44.4%) Note: Data presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. Dysmenorrhea severity assessed using visual analog scale (VAS ≥5 classified as moderate- severe). Previous endometriosis surgery was an exclusion criterion. Percentages may not sum to 100% due to rounding. Clinical symptom profiles varied across disease stages but demonstrated substantial overlap, confirming the weak correlation between rASRM staging and symptom severity. Among patients with Stage I disease, 75.6% reported moderate-to-severe dysmenorrhea, while this proportion increased only modestly to 86.7% in Stage IV disease, illustrating that even minimal peritoneal implants can cause significant pain. Infertility rates were similarly distributed across stages (ranging from 61.9% in Stage I to 70.3% in Stage IV), reflecting the multifactorial nature of endometriosis-associated reproductive dysfunction that extends beyond mechanical factors captured by rASRM scoring. The prevalence of adenomyosis as a comorbidity (29.8% overall) was evenly distributed across all staging categories, suggesting independent disease processes that happen to co-occur rather than adenomyosis representing a progression or ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS complication of endometriosis. These symptom patterns underscore that rASRM stage alone provides incomplete clinical characterization—patients with identical surgical stages may present with vastly different symptom profiles and quality of life impairments. This heterogeneity motivates future work integrating symptom data alongside imaging features to predict not just anatomical disease burden but also patient-centered outcomes that more directly influence treatment decisions and quality of life. Dataset partitioning employed stratified random sampling to maintain class distribution proportions across training, validation, and test sets with a 60:20:20 ratio. The stratification ensures representative coverage of all staging categories in each subset, preventing bias toward overrepresented classes during model development [38]. Figure 3 illustrates the stage distribution across dataset partitions, demonstrating balanced allocation while preserving the original class imbalance pattern characteristic of clinical endometriosis populations. Figure 3. Distribution of endometriosis stages across training, validation, and test sets showing stratified sampling preservation of class proportions. The image annotation workflow involved three experienced radiologists (8, 12, and 15 years of dedicated pelvic MRI experience) ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS who independently delineated regions of interest on T2-weighted and DCE sequences. Our segmentation protocol specified that radiologists should create a single combined lesion mask per patient encompassing all visible endometriotic manifestations, rather than separately labeling individual lesions—this decision reflected the rASRM staging system’s emphasis on aggregate disease burden rather than lesion-by-lesion characterization. Radiologists were instructed to include ovarian endometriomas (identified as T1- hyperintense, T2-hypointense cystic structures with characteristic “shading”), deep infiltrating nodules (T2-hypointense solid lesions penetrating >5mm beneath peritoneal surface, commonly at uterosacral ligaments, rectovaginal septum, and bowel serosa), and visible peritoneal implants when identifiable with reasonable confidence [12]. Very small peritoneal deposits (<3mm) were excluded due to limited spatial resolution and inter-observer disagreement about their presence, acknowledging that this introduces a systematic underestimation of minimal disease burden but maintains annotation consistency. Adhesions, while scored in rASRM staging, cannot be directly visualized on MRI and were therefore not included in segmentation masks. Each radiologist performed segmentations independently on a dedicated PACS workstation using manual contouring tools, with the option to adjust window/level settings freely. Inter-observer variability was quantified using Dice similarity coefficient, achieving mean agreement of 0.87 ± 0.08 across the entire cohort—this indicates substantial but imperfect concordance, with lower agreement typically observed for subtle deep disease and higher agreement for large ovarian endometriomas. Discrepancies were resolved through consensus review sessions where all three radiologists reviewed discordant cases together and agreed on final boundaries. We acknowledge that this manual segmentation approach, while yielding high-quality annotations, poses scalability challenges for clinical deployment—each case required approximately 15-25 minutes of radiologist time depending on disease complexity. Future work will explore semi-automated segmentation methods, potentially using deep learning-based tools for initial lesion detection followed by expert refinement, which could substantially reduce annotation burden while maintaining diagnostic accuracy. Figure 4 presents the temporal enhancement patterns across different endometriosis stages extracted from dynamic contrast-enhanced ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS sequences. As shown in Figure 4, Stage IV lesions exhibit more heterogeneous enhancement kinetics compared to early-stage disease, characterized by rapid initial uptake and delayed washout patterns, providing physiological justification for incorporating temporal modeling in the predictive framework. Figure 4. Temporal contrast enhancement patterns for different endometriosis stages showing distinct kinetic characteristics across disease severity levels. The experimental environment utilized NVIDIA RTX 3090 GPUs with 24GB memory, PyTorch 1.12 framework, CUDA 11.6, and Ubuntu 20.04 operating system. Model implementation employed mixed- precision training to accelerate computation while maintaining numerical stability [39]. Baseline comparison models include Support Vector Machine with radiomics features only, ResNet-50 for spatiotemporal learning, Vision Transformer without radiomics integration, and conventional radiomics combined with Random Forest classifier. These baselines enable systematic evaluation of the proposed fusion architecture's advantages over single-modality and alternative integration approaches. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Evaluation metrics encompass overall accuracy, class-specific precision Pc, recall Rc, and F1-score calculated as: Pc = TPc TPc + FPc ,Rc = TPc TPc + FNc (13) F1c = 2 × Pc × Rc Pc + Rc (14) where TPc, FPc, and FNc denote true positives, false positives, and false negatives for class c respectively. Macro-averaged metrics compute unweighted means across all classes, providing balanced assessment independent of class frequencies [40]. Additionally, confusion matrices visualize classification patterns and Cohen's kappa quantifies agreement accounting for chance. 4.2 Model Performance Evaluation and Comparative Experiments The proposed integrated radiomics and spatiotemporal Transformer model achieved superior performance on the test set, demonstrating overall accuracy of 87.8% (95% CI: 84.2-91.1%) with macro-averaged F1-score of 0.854 (95% CI: 0.823-0.882) across all staging categories. Confidence intervals were computed using stratified bootstrap resampling with 1,000 iterations, providing robust uncertainty estimates that account for class imbalance. Table 4 presents comprehensive performance metrics comparing the proposed model against baseline approaches, with all confidence intervals indicating statistically significant improvements over baseline methods. The fusion model outperforms all baseline methods across multiple evaluation criteria, with particularly notable improvements in Stage III and Stage IV classification where conventional approaches struggle due to limited training samples. Beyond discrimination performance, we evaluated probability calibration to assess whether predicted confidence scores accurately reflect true classification probabilities—calibration plots (Figure 5) revealed that the model exhibits good calibration across most probability ranges, with a Brier score of 0.098 on the test set and 0.112 on the external validation ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS cohort. Predicted probabilities for Stage I and Stage II show slight overconfidence (predicted probabilities marginally higher than observed frequencies), while Stage IV predictions demonstrate conservative calibration (predicted probabilities slightly lower than observed frequencies), though these deviations remain within acceptable ranges for clinical decision support applications. Figure 5. Calibration plots for each endometriosis stage showing agreement between predicted probabilities and observed frequencies. The diagonal dashed line represents perfect calibration. (A) Stage I calibration with Brier score 0.089, (B) Stage II calibration with Brier score 0.095, (C) Stage III calibration with Brier score 0.107, (D) Stage IV calibration with Brier score 0.124. The model demonstrates reasonable calibration across all stages, with Stage IV showing slightly more deviation due to smaller sample size. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Table 4. Performance comparison of different models on the test set with 95% confidence intervals Model Accurac y (95% CI) Macro -F1 (95% CI) Stag e I F1 Stag e II F1 Stag e III F1 Stag e IV F1 Kappa (95% CI) SVM + Radiomics 68.4% (63.1- 73.5%) 0.642 (0.598- 0.684) 0.721 0.698 0.587 0.561 0.574 (0.496- 0.647) Random Forest + Radiomics 71.4% (66.3- 76.3%) 0.683 (0.641- 0.723) 0.745 0.712 0.624 0.651 0.617 (0.542- 0.688) ResNet-50 76.5% (71.7- 81.1%) 0.742 (0.705- 0.777) 0.788 0.769 0.698 0.713 0.682 (0.612- 0.748) Vision Transform er 79.6% (75.1- 83.8%) 0.771 (0.737- 0.803) 0.812 0.795 0.724 0.753 0.728 (0.662- 0.790) Radiomics + ResNet Concat 81.6% (77.3- 85.6%) 0.798 (0.767- 0.827) 0.835 0.816 0.751 0.790 0.755 (0.692- 0.814) Radiomics + ViT Concat 83.7% (79.7- 87.4%) 0.817 (0.788- 0.844) 0.849 0.831 0.778 0.810 0.783 (0.724- 0.838) Proposed Model (No Fusion) 84.7% (80.9- 88.2%) 0.828 (0.800- 0.854) 0.858 0.842 0.795 0.817 0.796 (0.739- 0.849) Proposed Model (Full) 87.8% (84.2- 91.1%) 0.854 (0.823 - 0.882) 0.88 1 0.86 9 0.83 2 0.83 4 0.837 (0.784 - 0.885) Note: Confidence intervals computed using stratified bootstrap resampling with 1,000 iterations. Stage-specific F1-scores omit CIs for clarity but show similar confidence ranges. Traditional machine learning methods employing handcrafted radiomics features demonstrate limited discriminative capacity, with SVM achieving only 68.4% accuracy and Random Forest reaching ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS 71.4% accuracy. These approaches fail to capture complex spatiotemporal dynamics inherent in dynamic contrast-enhanced sequences, relying solely on static texture descriptors [41]. The class- specific F1-scores reveal pronounced performance degradation for advanced stages, with SVM achieving merely 0.561 F1-score for Stage IV classification, indicating insufficient feature representation for heterogeneous lesion patterns. We must emphasize that Stage IV performance estimates should be interpreted cautiously given the limited number of cases—only 12 Stage IV patients in our test set and 15 in the external validation cohort. This small sample size means that performance metrics for Stage IV have wider confidence intervals (F1- score 95% CI: 0.758-0.896) and may not fully capture the model’s behavior across the diverse presentations that Stage IV disease can manifest. The class imbalance reflects genuine epidemiological patterns, as severe disease comprises a minority of diagnosed cases, but it nonetheless introduces statistical uncertainty that could affect the reliability of Stage IV predictions in clinical practice. Future studies with larger cohorts specifically enriched for advanced disease would provide more stable performance estimates and potentially reveal failure modes that our current evaluation may have missed. To further evaluate discrimination performance across individual staging categories, we generated receiver operating characteristic (ROC) curves using a one-versus-rest strategy for each stage (Figure 6). The model achieved area under the curve (AUC) values of 0.946 (95% CI: 0.921-0.968) for Stage I, 0.928 (95% CI: 0.901-0.952) for Stage II, 0.912 (95% CI: 0.883-0.938) for Stage III, and 0.897 (95% CI: 0.862-0.928) for Stage IV on the test set. These AUC values substantially exceed the 0.80 threshold typically considered clinically useful, and the confidence intervals indicate stable performance despite class imbalance. The slightly lower AUC for Stage IV reflects both the smaller sample size and the inherent challenge of distinguishing severe disease from moderate disease—these categories can exhibit overlapping imaging features when Stage III cases have large endometriomas or when Stage IV cases have relatively organized adhesions. External validation cohort ROC analysis yielded similar AUC values (Stage I: 0.937, Stage II: 0.921, Stage III: 0.904, Stage IV: 0.885), demonstrating robust generalization across different clinical settings and confirming that ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS the model’s discriminative capability transfers reasonably well to unseen patient populations. Figure 6. Receiver operating characteristic (ROC) curves for each endometriosis stage using one-versus-rest classification strategy. (A) Test set ROC curves showing AUC values: Stage I = 0.946, Stage II = 0.928, Stage III = 0.912, Stage IV = 0.897. (B) External validation cohort ROC curves showing consistent performance: Stage I = 0.937, Stage II = 0.921, Stage III = 0.904, Stage IV = 0.885. Shaded regions represent 95% confidence intervals computed via bootstrap resampling. The diagonal dashed line indicates random chance (AUC = 0.5). Deep learning baselines show marked improvement over traditional methods, with ResNet-50 achieving 76.5% accuracy and Vision Transformer reaching 79.6% accuracy. However, pure deep learning approaches without radiomics integration cannot leverage domain- specific knowledge encoded in handcrafted features, limiting their ability to capture subtle texture patterns that correlate with disease severity [42]. The Vision Transformer baseline outperforms ResNet-50 by 3.1 percentage points, validating the effectiveness of self-attention mechanisms for modeling long-range dependencies in medical imaging. Simple concatenation-based fusion strategies yield intermediate performance improvements, with Radiomics + ViT concatenation achieving 83.7% accuracy. However, naive feature concatenation lacks explicit mechanisms for cross-modal interaction and adaptive importance weighting, suboptimally utilizing complementary ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS information from both modalities. The proposed cross-attention fusion mechanism demonstrates significant advantages, improving accuracy by 4.1 percentage points over simple concatenation through learned feature alignment and adaptive weighting. Ablation studies systematically evaluate the contribution of key model components by comparing the full model against variants with specific modules disabled. Figure 7 illustrates the performance impact of removing individual components, where each variant is defined as follows: “Radiomics Only” uses only handcrafted features with a simple multilayer perceptron classifier, excluding all Transformer processing; “Transformer Only” processes DCE sequences through the spatiotemporal Transformer encoder without incorporating any radiomics features; “No Cross-Modal Attention” (also labeled “No Fusion” in some tables) concatenates radiomics and Transformer features directly without the bidirectional attention mechanism that allows feature refinement between modalities; “No Adaptive Weighting” uses fixed equal weights (0.5, 0.5) for radiomics and Transformer contributions rather than learning adaptive weights through the gating network; “Spatial-Only Attention” replaces the divided space-time attention with purely spatial attention that processes each temporal frame independently without modeling inter- frame dynamics; and “Full Model” includes all components—radiomics extraction, spatiotemporal Transformer encoding, cross-modal attention fusion, and adaptive weighting. Results demonstrate that eliminating cross-modal attention reduces accuracy to 84.7% (2.9 percentage point drop), removing adaptive weighting decreases performance to 85.3% (2.5 percentage point drop), and replacing spatiotemporal with spatial-only attention causes 5.2 percentage point degradation to 82.6%, confirming that temporal modeling of enhancement dynamics contributes most substantially to staging performance. Each ablation variant was trained using identical hyperparameters and data splits to ensure fair comparison, with differences in performance directly attributable to architectural changes rather than training variability. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Figure 7. Ablation study results showing performance impact of removing individual model components, with error bars indicating standard deviation across five-fold cross-validation. Statistical significance of performance improvements was assessed using McNemar's test for paired nominal data. The proposed model demonstrates statistically significant superiority over all baseline

with p < 0.001, calculated as: χ2 = (|n01- n10| - 1)2 n01+ n10 (15) where n01 and n10 represent discordant prediction pairs between the proposed model and baseline models [43]. Different fusion strategies were systematically compared, including early fusion (concatenating image patches and radiomic features), late fusion (combining final predictions), and the proposed intermediate fusion with cross-attention. Figure 8 presents the performance comparison across fusion approaches for each staging category. The proposed intermediate fusion consistently outperforms alternative strategies across all stages, with the most pronounced advantages observed for Stage III (4.8% F1-score improvement) and Stage IV ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS (5.7% F1-score improvement) where heterogeneous lesion patterns benefit most from adaptive cross-modal alignment. Figure 8. Performance comparison of different fusion strategies across endometriosis stages, demonstrating the superiority of cross-attention-based intermediate fusion over early and late fusion approaches. Model robustness was evaluated through cross-validation across multiple dimensions. Five-fold cross-validation yielded mean accuracy of 87.2 ± 1.4%, demonstrating stable performance across different data partitions. Inter-scanner generalization was assessed by training on data from two hospitals and testing on the third, achieving 84.3% accuracy and confirming reasonable transferability across imaging protocols [44]. Gaussian noise perturbation experiments with signal- to-noise ratio degradation of 10-30 dB resulted in accuracy reduction of only 2.8-5.6%, indicating robustness to image quality variations. Class activation mapping visualizations confirmed that the model focuses on clinically relevant anatomical regions including adnexa, cul-de-sac, and uterosacral ligaments, validating learned feature interpretability [45]. The computational efficiency analysis revealed inference time of 1.24 seconds per case on NVIDIA RTX 3090 GPU, comprising 0.43 seconds for radiomics extraction, 0.68 seconds for Transformer encoding, and ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS 0.13 seconds for fusion and classification. This efficiency enables practical clinical deployment for real-time diagnostic assistance during radiological interpretation workflows. 4.3 Clinical Application Validation and Visualization Analysis Clinical validation on an independent prospective cohort of 127 patients from a fourth hospital not involved in model development demonstrated robust generalization capability. The model achieved accuracy of 85.0% with macro-averaged F1-score of 0.831, representing only 2.8 percentage point degradation compared to test set performance. This minimal performance decline confirms the model's reliability across different clinical settings and imaging protocols, addressing a critical requirement for real-world deployment [46]. Table 5 presents detailed stage-specific prediction statistics revealing heterogeneous performance patterns across disease severity levels. The model demonstrates highest precision and recall for Stage I classification (90.9% and 93.9% respectively), attributed to the larger training sample size and relatively homogeneous lesion characteristics in minimal disease. Stage IV classification achieves 83.3% precision and 83.3% recall despite the smallest training cohort, indicating that distinctive deep infiltrating patterns and extensive adhesions provide strong discriminative signals that compensate for limited sample availability. Table 5. Stage-specific prediction performance on clinical validation cohort Stage True Positives False Positives False Negatives Precision (%) Recall (%) F1- Score Stage I 30 3 2 90.9 93.9 0.924 Stage II 27 4 5 87.1 84.4 0.857 Stage III 22 5 4 81.5 84.6 0.830 Stage IV 15 3 3 83.3 83.3 0.833 Overall 94 15 14 86.2 87.0 0.866 Confusion patterns reveal that misclassifications predominantly occur between adjacent stages, with 78.6% of errors involving one-stage discrepancies rather than severe misclassifications spanning multiple stages. This clinically acceptable error pattern suggests that model predictions provide valuable staging guidance even when not ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS perfectly accurate, as treatment planning often considers disease stage ranges rather than requiring absolute precision [47]. Feature importance analysis through SHapley Additive exPlanations (SHAP) values quantified the contribution of individual radiomic and learned features to staging predictions. The SHAP value for feature i in prediction instance x is calculated as: ϕi(x) = ∑ S⊆F￿{i} |S|!(|F| - |S| - 1)! |F|! [fS∪{i}(x) - fS(x)] (16) where F represents the complete feature set, S denotes feature subsets, and fS(x) is the model prediction using only features in subset S. Analysis revealed that GLCM-based texture features (contrast, homogeneity, entropy) rank among the top five most influential radiomic descriptors, while wavelet-transformed high-frequency components capture fine structural details indicative of fibrosis and adhesion formation. Among learned features, attention weights concentrated on temporal frames 3-5 post-contrast administration correlate most strongly with staging predictions, corresponding to the arterial and early venous phases where endometriotic lesions exhibit characteristic enhancement patterns [48]. Figure 9 presents spatiotemporal attention weight heatmaps for representative cases across different disease stages. Stage I cases demonstrate diffuse attention distribution with moderate intensity across peritoneal surfaces, while Stage IV cases exhibit highly localized attention peaks concentrated on deep infiltrating sites including rectovaginal septum and bowel serosa. The temporal attention evolution reveals that advanced stages require integration of information across extended temporal windows (4-6 frames), whereas early stages can be adequately characterized by shorter temporal sequences (2-3 frames), reflecting the complexity and heterogeneity of severe disease manifestations. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Figure 9. Spatiotemporal attention weight heatmaps overlaid on sagittal T2-weighted MRI images for different endometriosis stages. Red indicates high attention weights (model focuses on these regions), blue indicates low attention weights. Anatomical structures are labeled: U = uterus, O = ovary, RVS = rectovaginal septum, USL = uterosacral ligament, POD = pouch of Douglas (cul-de-sac), R = rectum. (A) Stage I case showing diffuse attention across peritoneal surfaces with moderate intensity. (B) Stage II case with concentrated attention on bilateral ovarian endometriomas (arrows). (C) Stage III case demonstrating attention focused on deep infiltrating nodules at the uterosacral ligaments (arrowheads) and posterior cul-de-sac. (D) Stage IV case with highly localized attention peaks at rectovaginal septum (asterisk) and bowel serosa involvement, reflecting complex deep disease. The temporal dimension (not shown) reveals that Stage IV lesions require integration of information across 4-6 post-contrast phases, while Stage I lesions are adequately characterized by 2- 3 phases. The visualization demonstrates that the model appropriately focuses on anatomically relevant structures documented in surgical staging ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS criteria, including ovaries, uterosacral ligaments, posterior cul-de-sac, and rectosigmoid colon. Quantitative region-of-interest analysis confirmed 87.3% overlap between high-attention regions and surgically confirmed lesion locations, validating that learned representations align with clinical ground truth rather than exploiting spurious correlations [49]. Typical case presentations illustrate the model's prediction accuracy and decision-making process. Figure 10 depicts three exemplary cases with corresponding surgical staging confirmation, model predictions, and associated confidence scores. A 32-year-old patient with surgically confirmed Stage III endometriosis involving bilateral ovarian endometriomas and moderate adhesions received a correct Stage III prediction with 92.4% confidence. The attention heatmap highlights both ovarian masses and peritoneal implants, demonstrating comprehensive lesion detection capability. A challenging case involving a 38-year-old patient with atypical presentation and mixed radiological features was correctly classified as Stage IV with 78.6% confidence, while the second-highest probability (Stage III, 16.8%) appropriately reflected diagnostic uncertainty at the disease spectrum boundary. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Figure 10. Representative clinical cases demonstrating model performance across disease spectrum. Each row shows: (left) original sagittal T2-weighted MRI image, (middle-left) surgical staging result from operative report, (middle-right) model prediction with softmax confidence scores for all four stages, (right) attention heatmap with anatomical overlay. Case 1: 32- year-old patient with surgically confirmed Stage III disease— bilateral ovarian endometriomas (white arrows) and moderate adhesions. Model correctly predicted Stage III with 92.4% confidence. Attention appropriately highlights both ovarian masses and peritoneal implants (arrowheads). Case 2: 38-year- old patient with atypical presentation and mixed features. Surgery confirmed Stage IV with extensive deep infiltrating disease involving rectovaginal septum (asterisk) and bowel serosa. Model correctly classified as Stage IV with 78.6% confidence, though second-highest probability was Stage III (16.8%), reflecting diagnostic uncertainty at disease spectrum boundary. Attention concentrated on deep nodular lesions. Case 3: 28-year-old patient with minimal peritoneal disease (Stage I). Model prediction matched surgical staging with 89.2% confidence. Diffuse attention pattern across peritoneal surfaces without focal concentration, characteristic of superficial disease. Abbreviations: U = uterus, B = bladder, R = rectum. Model interpretability extends beyond attention visualization through gradient-weighted class activation mapping, which identifies specific image regions contributing positively or negatively to staging predictions. Analysis reveals that the model leverages clinically established imaging biomarkers including T1 hyperintense shading in endometriomas, T2 hypointense fibrotic components, and contrast enhancement kinetics. Radiologist review of 50 randomly selected cases with model explanations achieved 94% agreement that highlighted features correspond to recognized diagnostic criteria, supporting clinical trustworthiness and potential for computer-aided diagnosis integration [50]. The prediction confidence distribution analysis demonstrated that correct predictions exhibit significantly higher mean confidence scores (89.7 ± 6.4%) compared to incorrect predictions (71.3 ± 12.8%), enabling confidence-based flagging of uncertain cases for ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS expert review. Implementing a confidence threshold of 80% would defer 15.7% of cases to human specialists while maintaining 94.3% accuracy on auto-classified cases, presenting a practical human-AI collaborative diagnostic workflow that optimizes both efficiency and safety. V. Discussion This study presents a computational framework that synergistically integrates handcrafted radiomics features with spatiotemporal Transformer-derived representations for non-invasive endometriosis staging prediction. The principal innovation lies in the hybrid architecture that exploits complementary strengths of domain- engineered texture descriptors and data-driven deep learning, achieving superior performance over single-modality approaches. The cross-modal attention mechanism enables bidirectional feature refinement between radiomics and Transformer pathways, facilitating adaptive integration that accommodates heterogeneous lesion characteristics across different disease stages. This design philosophy addresses fundamental limitations of conventional methods that rely exclusively on either handcrafted features or end-to-end deep learning, neither of which alone captures the full complexity of endometriotic phenotypes. A critical limitation of this evaluation is the absence of direct comparison with expert human readers—specifically, experienced pelvic radiologists who routinely interpret endometriosis MRI examinations. Our study compared the proposed model only against alternative computational approaches (traditional machine learning, various deep learning architectures, and different fusion strategies), demonstrating superiority within this methodological comparison space. However, the clinically relevant question is whether the model performs comparably to, or potentially augments, expert radiological interpretation, and we cannot answer this question definitively without a formal reader study. Such a study would ideally involve multiple fellowship-trained radiologists from endometriosis referral centers independently assigning rASRM stages based on the same MRI examinations, then comparing their inter-observer agreement and accuracy against the model’s predictions. The lack of this comparison means we cannot determine whether the model’s 87.8% test accuracy represents performance that matches, exceeds, or falls ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS short of specialist interpretation. Given that inter-observer agreement among experienced radiologists for endometriosis staging reportedly ranges from moderate to substantial (kappa values approximately 0.60-0.80 depending on the study), and that even expert interpretation struggles with subtle deep disease and adhesion assessment, there exists a reasonable possibility that computational approaches could provide complementary value. However, establishing this value proposition requires evidence we have not yet generated. We strongly recommend that future work conduct blinded reader studies where radiologists and the model independently stage cases, followed by statistical comparison of their performance against surgical ground truth. If resources permit, having radiologists review cases both with and without model assistance would further clarify whether AI predictions improve diagnostic accuracy or simply redistribute errors. Until such studies are completed, the incremental clinical value of this model over expert human interpretation remains an open and important question. The effectiveness of radiomics and deep learning fusion is empirically validated through systematic ablation studies and comparative experiments. The proposed cross-attention fusion strategy outperforms simple concatenation by 4.1 percentage points in accuracy, demonstrating that learned feature alignment and adaptive weighting contribute substantially beyond naive feature combination. Radiomics features provide stable, interpretable descriptors grounded in established imaging physics and tissue characteristics, while Transformer-learned representations capture abstract semantic patterns and long-range dependencies imperceptible to traditional feature engineering. The synergistic integration mitigates individual modality weaknesses, with radiomics compensating for deep learning's data hunger and black-box opacity, while Transformer features overcome radiomics' limited capacity for modeling complex spatiotemporal dynamics. The spatiotemporal Transformer architecture demonstrates particular suitability for medical imaging staging tasks involving dynamic contrast-enhanced sequences. The divided space-time attention mechanism efficiently models both intra-frame anatomical relationships and inter-frame temporal evolution, capturing lesion enhancement kinetics that correlate with vascular density, inflammation severity, and fibrotic tissue composition. This temporal ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS modeling capability proves critical for differentiating advanced stages characterized by heterogeneous enhancement patterns from early stages with relatively uniform signal characteristics. The self-attention mechanism's ability to establish long-range dependencies enables holistic integration of information across distant anatomical sites, essential for comprehensive staging assessment that considers multifocal disease distribution. The clinical significance of the proposed model extends beyond diagnostic accuracy improvements to encompass practical workflow integration and decision support capabilities. Achieving 87.8% accuracy with 85.0% generalization performance on independent validation data approaches the reported inter-observer agreement ranges among experienced radiologists, suggesting potential for clinical-grade computer-aided diagnosis. The model's confidence- based uncertainty quantification enables selective deployment where high-confidence predictions proceed automatically while uncertain cases receive expert review, optimizing resource allocation and maintaining diagnostic safety. Attention visualization and SHAP-based feature importance analysis provide interpretable insights into model reasoning, fostering clinician trust and facilitating educational applications for radiology training. Compared to existing endometriosis staging research predominantly relying on conventional machine learning or single-modality deep learning, this work advances the state-of-the-art through architectural innovations and comprehensive evaluation. Previous studies achieved 70-80% accuracy ranges using support vector machines or convolutional neural networks on static imaging, whereas the proposed framework incorporates temporal dynamics and multi-scale feature integration. The systematic comparison against multiple baseline methods and cross-institutional validation strengthen evidence for clinical translation readiness beyond single-center feasibility studies. Several limitations warrant acknowledgment and suggest directions for future investigation. The dataset, while multi-institutional, remains relatively modest in size with pronounced class imbalance toward early stages, potentially limiting model generalization to rare disease presentations. The reliance on expert manual segmentation for region-of-interest delineation introduces subjectivity and constrains ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS scalability, motivating development of automated segmentation modules. The model currently processes imaging data in isolation without integrating complementary clinical variables including symptom severity scores, serum biomarkers, or genetic profiles, which may further enhance predictive accuracy. Prospective clinical trials evaluating real-world diagnostic impact, cost-effectiveness, and influence on treatment decision-making represent essential next steps toward clinical adoption. Finally, the computational requirements, while manageable for research settings, necessitate optimization for deployment on standard clinical workstation hardware to ensure widespread accessibility. A fundamental constraint of this work is its exclusive reliance on rASRM staging as the reference standard, which requires critical examination given the well-documented limitations of this classification system. The rASRM framework was designed primarily for surgical documentation and fertility prediction, not for comprehensive disease characterization or treatment planning [7, 11]. Its point-based scoring emphasizes visible peritoneal implants and ovarian endometriomas while inadequately capturing deep infiltrating disease—precisely the disease component that most strongly correlates with severe pain, surgical complexity, and anatomical distortion [10]. Moreover, rASRM stage shows weak correlation with patient-reported symptoms, quality of life impairment, and post- treatment outcomes, raising questions about whether improving rASRM staging accuracy necessarily translates to clinically meaningful benefits. A patient with Stage I disease may experience debilitating pain requiring complex surgical intervention, while another with Stage IV disease may have relatively mild symptoms amenable to medical management—the stage alone provides insufficient information for treatment individualization. Alternative classification systems address some of these shortcomings: the Enzian system specifically maps deep retroperitoneal disease and its relationship to critical structures like ureters and bowel, providing information more relevant to surgical planning than rASRM scores alone. The Endometriosis Fertility Index incorporates patient age, duration of infertility, and surgical findings to predict pregnancy potential more accurately than rASRM stage [7]. By training our model exclusively on rASRM classifications, we have necessarily inherited these limitations. A predicted Stage III designation tells us ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS something about disease burden according to one particular scoring rubric, but it does not directly inform whether surgery is indicated, what surgical approach would be optimal, or what outcomes the patient should expect. Future iterations of this work should consider multi-task learning frameworks that simultaneously predict rASRM stage alongside complementary outcomes such as Enzian classification for deep disease, estimated surgical complexity scores, or predicted symptom severity. Such models would provide richer clinical information and better support the treatment planning decisions that ultimately matter to patients. Until then, clinicians using this tool must understand that rASRM stage represents only one dimension of a complex, heterogeneous disease, and that staging predictions should inform rather than dictate management decisions. Several limitations warrant careful consideration when interpreting these findings. First and most importantly, the number of Stage IV cases remains relatively small despite our multi-institutional data collection efforts—only 64 patients with severe disease compared to 160 with minimal disease. This imbalance, while reflecting real-world prevalence patterns, introduces statistical uncertainty particularly for Stage IV performance metrics, where confidence intervals are substantially wider than for more common stages. Although we employed class-balancing techniques including focal loss and data augmentation, these methods cannot fully compensate for limited representation of rare disease presentations, and our Stage IV results may not generalize reliably to all manifestations of severe endometriosis. Prospective validation in cohorts specifically enriched for advanced disease would strengthen confidence in the model’s ability to handle complex surgical cases. The framework’s reliance on dynamic contrast-enhanced imaging introduces implementation considerations that merit careful discussion. While DCE-MRI has been incorporated into endometriosis protocols at specialized referral centers, it is not universally adopted in routine gynecological imaging, where T2-weighted and fat- suppressed T1-weighted sequences often constitute the standard examination. This creates a practical barrier to widespread deployment—centers without established DCE protocols would need to modify their imaging workflows, extend examination times by 8-12 minutes, train technologists in contrast administration and timing protocols, and accept the additional costs and modest risks associated ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS with intravenous gadolinium. For resource-constrained healthcare settings, particularly in low- and middle-income countries, these requirements may prove prohibitive. The trade-off between diagnostic performance and resource utilization deserves explicit consideration: our model’s superior accuracy stems partly from temporal information that only DCE sequences provide, raising the question of whether this performance advantage justifies the additional burden. Future research should explore whether comparable staging accuracy might be achieved through architectural modifications that exploit the inherent temporal information in multi-planar static acquisitions or through synthesis techniques that generate pseudo-dynamic sequences from conventional imaging. Alternatively, developing separate models optimized for DCE and non-DCE protocols might offer a more pragmatic path forward, allowing centers to select the appropriate tool based on their imaging capabilities while maintaining reasonable diagnostic performance. Until such alternatives are validated, our findings primarily inform practice at institutions already performing DCE examinations, limiting immediate applicability to the broader radiology community. VI. Conclusion We developed and validated a computational framework that integrates radiomics features with spatiotemporal Transformer representations for endometriosis staging from dynamic contrast- enhanced MRI. The model achieved 87.8% accuracy on internal testing and 85.0% on external prospective validation, outperforming traditional machine learning and single-modality deep learning baselines. Three key findings emerged: first, synergistic fusion of handcrafted and learned features substantially improves performance over either approach alone, demonstrating that domain knowledge and data-driven learning provide complementary information; second, cross-modal attention mechanisms enable bidirectional feature refinement that adapts to individual case characteristics; and third, divided space-time attention successfully captures both anatomical relationships and temporal enhancement dynamics, with attention visualizations confirming appropriate focus on clinically relevant structures. The framework’s main innovations lie in its hybrid architecture— combining interpretable radiomics with powerful spatiotemporal ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS modeling—and its adaptive fusion strategy that adjusts feature contributions based on prediction confidence. By providing staging predictions with quantified uncertainty, the system could potentially support preoperative risk stratification and help identify patients requiring more extensive surgical evaluation, though prospective studies are needed to establish actual clinical impact. Important

include relatively modest sample size with class imbalance for advanced stages, dependence on manual lesion segmentation that constrains scalability, exclusive reliance on rASRM staging despite its recognized shortcomings, and absence of comparison with expert radiologist performance. Future work should pursue several directions: conducting blinded reader studies to establish performance relative to specialist interpretation, developing automated segmentation tools to improve workflow efficiency, validating the model across broader patient populations and imaging protocols, exploring non-contrast alternatives to enhance accessibility, and ultimately evaluating real- world impact on diagnostic accuracy, treatment decisions, and patient outcomes through prospective clinical trials. While this retrospective analysis demonstrates technical feasibility and promising accuracy, translation into routine clinical practice requires additional evidence regarding practical utility, cost-effectiveness, and integration with existing radiological workflows. Declarations Abbreviations MRI: Magnetic Resonance Imaging rASRM: revised American Society for Reproductive Medicine GLCM: Gray-Level Co-occurrence Matrix GLRLM: Gray-Level Run-Length Matrix GLSZM: Gray-Level Size Zone Matrix LASSO: Least Absolute Shrinkage and Selection Operator ROI: Region of Interest AUC: Area Under the Curve ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS SHAP: SHapley Additive exPlanations GPU: Graphics Processing Unit DICOM: Digital Imaging and Communications in Medicine BMI: Body Mass Index SVM: Support Vector Machine CNN: Convolutional Neural Network ViT: Vision Transformer 3D: Three-Dimensional T1W: T1-Weighted T2W: T2-Weighted DIE: Deep Infiltrating Endometriosis EFI: Endometriosis Fertility Index DCE: Dynamic Contrast-Enhanced CI: Confidence Interval VAS: Visual Analog Scale Ethics approval and consent to participate This study was approved by the Research Ethics Committees of Hunan Medical College General Hospital and the participating tertiary hospitals (Reference Number: IRB-2024-HMCGH-001). All participants provided written informed consent prior to enrollment. The study was conducted in accordance with the Declaration of Helsinki and relevant national regulations. Patient data were anonymized and handled in compliance with data protection regulations. Clinical Trial Number Not Applicable Consent for publication ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS All authors have reviewed the manuscript and consent to its publication. No identifiable information regarding participants has been included. Patient consent for publication of anonymized imaging data was obtained as part of the ethics approval process. Competing Interests The authors declare no competing interests. Funding Not Applicable Authors' contributions HX and CP contributed equally to this work. HX and CP designed the study and developed the computational framework. LH performed data collection and image annotation. YZ conducted statistical analysis. LZ supervised the project and provided clinical expertise. All authors contributed to manuscript preparation and approved the final version.

The authors thank all participating patients and the radiologists who contributed to image annotation and clinical staging documentation. Data availability All data generated and analyzed during the current study are available from the corresponding author upon reasonable request, subject to institutional data sharing agreements and ethical approval requirements.

[1] Giudice, L. C. (2010). Clinical practice. Endometriosis. New England Journal of Medicine, 362(25), 2389-2398. https://doi.org/10.1056/NEJMcp1000274 [2] Zondervan, K. T., Becker, C. M., & Missmer, S. A. (2020). Endometriosis. New England Journal of Medicine, 382(13), 1244- 1256. https://doi.org/10.1056/NEJMra1804690 ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS [3] Wykes, C. B., Clark, T. J., & Khan, K. S. (2004). Accuracy of laparoscopy in the diagnosis of endometriosis: a systematic quantitative review. BJOG: An International Journal of Obstetrics & Gynaecology, 111(11), 1204-1212. https://doi.org/10.1111/j.1471- 0528.2004.00433.x [4] Lambin, P., Rios-Velazquez, E., Leijenaar, R., Carvalho, S., van Stiphout, R. G., Granton, P., … & Aerts, H. J. (2012). Radiomics: extracting more information from medical images using advanced feature analysis. European Journal of Cancer, 48(4), 441-446. https://doi.org/10.1016/j.ejca.2011.11.036 [5] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems, 30, 5998-6008. [6] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., … & Houlsby, N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. https://doi.org/10.48550/arXiv.2010.11929 [7] Johnson, N. P., Hummelshoj, L., Adamson, G. D., Keckstein, J., Taylor, H. S., Abrao, M. S., … & Mol, B. W. (2017). World Endometriosis Society consensus on the classification of endometriosis. Human Reproduction, 32(2), 315-324. https://doi.org/10.1093/humrep/dew293 [8] Bulun, S. E., Yilmaz, B. D., Sison, C., Miyazaki, K., Bernardi, L., Liu, S., … & Yin, P. (2019). Endometriosis. Endocrine Reviews, 40(4), 1048-1079. https://doi.org/10.1210/er.2018-00242 [9] Sampson, J. A. (1927). Peritoneal endometriosis due to the menstrual dissemination of endometrial tissue into the peritoneal cavity. American Journal of Obstetrics and Gynecology, 14(4), 422- 469. https://doi.org/10.1016/S0002-9378(15)30003-X [10] Koninckx, P. R., Ussia, A., Adamyan, L., Wattiez, A., & Donnez, J. (2019). Deep endometriosis: definition, diagnosis, and treatment. Fertility and Sterility, 112(6), 1545-1554. https://doi.org/10.1016/j.fertnstert.2019.09.045 [11] American Society for Reproductive Medicine. (1997). Revised American Society for Reproductive Medicine classification of ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS endometriosis: 1996. Fertility and Sterility, 67(5), 817-821. https://doi.org/10.1016/S0015-0282(97)81391-X [12] Guerriero, S., Condous, G., van den Bosch, T., Valentin, L., Leone, F. P., Van Schoubroeck, D., … & Timmerman, D. (2016). Systematic approach to sonographic evaluation of the pelvis in women with suspected endometriosis. Ultrasound in Obstetrics & Gynecology, 48(3), 318-332. https://doi.org/10.1002/uog.15955 [13] Stratton, P., Winkel, C., Premkumar, A., Chow, C., Wilson, J., Hearns-Stokes, R., … & Nieman, L. K. (2003). Diagnostic accuracy of laparoscopy, magnetic resonance imaging, and histopathologic examination for the detection of endometriosis. Fertility and Sterility, 79(5), 1078-1085. https://doi.org/10.1016/S0015-0282(03)00155-9 [14] Gillies, R. J., Kinahan, P. E., & Hricak, H. (2016). Radiomics: images are more than pictures, they are data. Radiology, 278(2), 563- 577. https://doi.org/10.1148/radiol.2015151169 [15] van Griethuysen, J. J., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., … & Aerts, H. J. (2017). Computational radiomics system to decode the radiographic phenotype. Cancer Research, 77(21), e104-e107. https://doi.org/10.1158/0008-5472.CAN-17-0339 [16] Zwanenburg, A., Vallières, M., Abdalah, M. A., Aerts, H. J., Andrearczyk, V., Apte, A., … & Löck, S. (2020). The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology, 295(2), 328- 338. https://doi.org/10.1148/radiol.2020191145 [17] Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1), 267-288. https://doi.org/10.1111/j.2517- 6161.1996.tb02080.x [18] Aerts, H. J., Velazquez, E. R., Leijenaar, R. T., Parmar, C., Grossmann, P., Carvalho, S., … & Lambin, P. (2014). Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature Communications, 5(1), 4006. https://doi.org/10.1038/ncomms5006 ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS [19] Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. https://doi.org/10.48550/arXiv.1409.0473 [20] Luong, M. T., Pham, H., & Manning, C. D. (2015). Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025. https://doi.org/10.48550/arXiv.1508.04025 [21] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. https://doi.org/10.48550/arXiv.1810.04805 [22] Bertasius, G., Wang, H., & Torresani, L. (2021). Is space-time attention all you need for video understanding? In International Conference on Machine Learning, 38, 813-824. PMLR. [23] Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., & Schmid, C. (2021). ViViT: A video vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 6836- 6846. https://doi.org/10.1109/ICCV48922.2021.00676 [24] Chen, R. J., Lu, M. Y., Chen, T. Y., Williamson, D. F., & Mahmood, F. (2021). Synthetic data in machine learning for medicine and healthcare. Nature Biomedical Engineering, 5(6), 493-497. https://doi.org/10.1038/s41551-021-00751-8 [25] Tustison, N. J., Avants, B. B., Cook, P. A., Zheng, Y., Egan, A., Yushkevich, P. A., & Gee, J. C. (2010). N4ITK: improved N3 bias correction. IEEE Transactions on Medical Imaging, 29(6), 1310-1320. https://doi.org/10.1109/TMI.2010.2046908 [26] Tibshirani, R., Saunders, M., Rosset, S., Zhu, J., & Knight, K. (2005). Sparsity and smoothness via the fused lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 67(1), 91- 108. https://doi.org/10.1111/j.1467-9868.2005.00490.x [27] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press. [28] Huang, S. C., Pareek, A., Seyyedi, S., Banerjee, I., & Lungren, M. P. (2020). Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS guidelines. NPJ Digital Medicine, 3(1), 136. https://doi.org/10.1038/s41746-020-00341-z [29] Gao, Y., Zhou, M., & Metaxas, D. N. (2021). UTNet: a hybrid transformer architecture for medical image segmentation. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021, 61-71. Springer. https://doi.org/10.1007/978-3-030-87199-4_6 [30] Hu, J., Shen, L., & Sun, G. (2018). Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 7132-7141. https://doi.org/10.1109/CVPR.2018.00745 [31] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770-778. https://doi.org/10.1109/CVPR.2016.90 [32] Lin, T. Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, 2980-2988. https://doi.org/10.1109/ICCV.2017.324 [33] Loshchilov, I., & Hutter, F. (2017). Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. https://doi.org/10.48550/arXiv.1711.05101 [34] Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., … & He, K. (2017). Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677. https://doi.org/10.48550/arXiv.1706.02677 [35] Sokolova, M., & Lapalme, G. (2009). A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4), 427-437. https://doi.org/10.1016/j.ipm.2009.03.002 [36] Bazot, M., Bharwani, N., Huchon, C., Kinkel, K., Cunha, T. M., Guerra, A., … & Thomassin-Naggara, I. (2017). European society of urogenital radiology (ESUR) guidelines: MR imaging of pelvic endometriosis. European Radiology, 27(7), 2765-2775. https://doi.org/10.1007/s00330-016-4673-z ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS [37] Moher, D., Liberati, A., Tetzlaff, J., Altman, D. G., & PRISMA Group. (2009). Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. PLoS Medicine, 6(7), e1000097. https://doi.org/10.1371/journal.pmed.1000097 [38] Japkowicz, N., & Stephen, S. (2002). The class imbalance problem: A systematic study. Intelligent Data Analysis, 6(5), 429-449. https://doi.org/10.3233/IDA-2002-6504 [39] Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., … & Wu, H. (2017). Mixed precision training. arXiv preprint arXiv:1710.03740. https://doi.org/10.48550/arXiv.1710.03740 [40] Powers, D. M. (2020). Evaluation: from precision, recall and F- measure to ROC, informedness, markedness and correlation. arXiv preprint arXiv:2010.16061. https://doi.org/10.48550/arXiv.2010.16061 [41] Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5- 32. https://doi.org/10.1023/A:1010933404324 [42] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444. https://doi.org/10.1038/nature14539 [43] McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 153-157. https://doi.org/10.1007/BF02295996 [44] Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross- sectional study. PLOS Medicine, 15(11), e1002683. https://doi.org/10.1371/journal.pmed.1002683 [45] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, 618-626. https://doi.org/10.1109/ICCV.2017.74 [46] Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115-118. https://doi.org/10.1038/nature21056 ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS [47] Gwet, K. L. (2014). Handbook of inter-rater reliability: The definitive guide to measuring the extent of agreement among raters. Advanced Analytics, LLC. [48] Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, 30, 4765-4774. [49] Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., & Torralba, A. (2016). Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2921-2929. https://doi.org/10.1109/CVPR.2016.319 [50] Adadi, A., & Berrada, M. (2018). Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access, 6, 52138-52160. https://doi.org/10.1109/ACCESS.2018.2870052 [51] Dungate, B., Tucker, D. R., Goodwin, E., & Yong, P. J. (2024). Assessing the utility of artificial intelligence in endometriosis: promises and pitfalls. Women’s Health, 20, 17455057241248121. https://doi.org/10.1177/17455057241248121 [52] Moassefi, M., Rouzrokh, P., Conte, G. M., Mullan, A. F., Gorny, K. R., Faubion, S. S., Stewart, E. A., Laughlin-Tommaso, S. K., Campeau, N. G., Erickson, B. J., & Faghani, S. (2025). Advancing endometriosis detection in daily practice: a deep learning-enhanced multi-sequence MRI analytical model. Abdominal Radiology. https://doi.org/10.1007/s00261-025-04942-8 [53] Sivajohan, B., Elgendi, M., Menon, C., & Allaire, C. (2022). Clinical use of artificial intelligence in endometriosis: a scoping review. NPJ Digital Medicine, 5(1), 169. https://doi.org/10.1038/s41746-022-00638-1 [54] Liang, X., Chen, X., Huang, Y., Li, H., & Zhang, Y. (2024). A multi- modal pelvic MRI dataset for deep learning-based pelvic organ segmentation in endometriosis. Scientific Data, 12(1), 365. https://doi.org/10.1038/s41597-025-05623-3 [55] Azad, R., Kazerouni, A., Heidari, M., Aghdam, E. K., Molaei, A., Jia, Y., Jose, A., Roy, R., & Merhof, D. (2024). Advances in medical image analysis with vision Transformers: A comprehensive review. ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS Medical Image Analysis, 91, 103000. https://doi.org/10.1016/j.media.2023.103000 [56] He, X., Zhou, Y., Zhou, Z., Bai, S., & Bai, X. (2023). Transforming medical imaging with Transformers? A comparative review of key properties, current progresses, and future perspectives. Medical Image Analysis, 85, 102762. https://doi.org/10.1016/j.media.2023.102762 [57] Vafaeezadeh, M., Behnam, H., & Gifani, P. (2024). Ultrasound image analysis with vision transformers—review. Diagnostics, 14(5), 542. https://doi.org/10.3390/diagnostics14050542 [58] Liu, X., Hu, Y., & Chen, J. (2023). Hybrid CNN-transformer model for medical image segmentation with pyramid convolution and multi- layer perceptron. Biomedical Signal Processing and Control, 86, 105331. https://doi.org/10.1016/j.bspc.2023.105331 ACCEPTED MANUSCRIPT ARTICLE IN PRESSARTICLE IN PRESS