{"paper_id":"76e885c3-3e6e-4d1e-a76f-99d93b07addd","body_text":"Acadlore Transactions on AI and Machine Learning\nhttps://www.acadlore.com/journals/ATAIML\nDecision-Level Multimodal Fusion for Non-Invasive Diagnosis of\nEndometriosis: Strategies, Calibration, and Net Clinical Benefit\nOluwayemisi B. Fatade1*\n , Oyebimpe F. Ajiboye2\n , Funmilayo A. Sanusi3\n , Kikelomo I. Okesola3\n ,\nGrace C. Okorie3\n , Goodness O. Opateye1\n , Oluwasefunmi B. Famodimu1\n1 Department of Computer Science, Babcock University, 121003 Ilishan Remo, Nigeria\n2 Wosler Diagnostics, T2H 2B2 Calgary, Canada\n3 Department of Software Engineering, Babcock University, 121003 Ilishan Remo, Nigeria\n* Correspondence: Oluwayemisi B. Fatade (fatadeo@babcock.edu.ng)\nReceived: 11-05-2025 Revised: 12-23-2025 Accepted: 01-08-2026\nCitation: O. B. Fatade, O. F. Ajiboye, F. A. Sanusi, K. I. Okesola, G. C. Okorie, G. O. Opateye, and O. B.\nFamodimu, “Decision-level multimodal fusion for non-invasive diagnosis of endometriosis: Strategies, calibration,\nand net clinical benefit,” Acadlore Trans. Mach. Learn., vol. 5, no. 1, pp. 11–19, 2026. https://doi.org/10.56578/a\ntaiml050102.\n 2026 by the author(s). Licensee Acadlore Publishing Services Limited, Hong Kong. This article can be downloaded for free, and\nreused and quoted with a citation of the original published version, under the CC BY 4.0 license.\nAbstract: Endometriosis remains underdiagnosed due to reliance on invasive laparoscopy. Artificial Intelligence\n(AI) using either imaging or structured clinical data have shown promise, but single modality approaches face\nlimitations in sensitivity, calibration, and clinical reliability. This work seeks to evaluate whether decision-level\nmultimodal fusion of Magnetic Resonance Imaging (MRI)-based and clinical data-based AI systems improves\ndiagnostic performance, calibration, and net clinical benefit, compared with single-modality models. Two previously\nvalidated models were combined with retrospective data from 1,208 patients with suspected endometriosis: a Dual\nU-Net trained on pelvic MRI with Gradient-weighted Class Activation Mapping (Grad-CAM) interpretability and a\ndense neural network trained on structured clinical features with SHapley Additive exPlanations (SHAP). This study\ntested weighted averaging, stacking via logistic regression, and confidence-gating. Performance was assessed using\naccuracy, precision, recall, F1-score, and area under the curve (AUC). Calibration was evaluated using the Brier\nscore, expected calibration error (ECE), and reliability diagrams. Clinical utility was quantified with decision curve\nanalysis (DCA). Statistical significance was tested with McNemar’s test for accuracy and DeLong’s test for AUC.\nMultimodal fusion outperformed both single modality models. Weighted averaging accuracy was 0.89, precision was\n0.89, recall was 0.87, and F1-score was 0.86, thus improving on either modality alone. Stacking further enhanced\ncalibration (ECE reduction from 0.8 to 0.04) and yielded higher net benefit across clinically relevant probability\nthresholds (20 to 60%). DCA indicated fusion would avoid 12 to 18 unnecessary surgical investigations per 100\npatients, compared with single modality strategies. Confidence-gating maintained performance under simulated\ndistribution shifts to support robustness. Decision-level multimodal fusion enhanced non-invasive diagnosis of\nendometriosis by improving accuracy, calibration, and clinical utility. These results demonstrated the value of\nintegrative AI gynecological care and justify prospective validation in real-world clinical settings.\nKeywords: Endometriosis; Multimodal fusion; Gradient-weighted Class Activation Mapping; Neural U-Net\n1 Introduction\nEndometriosis is a chronic gynecological condition affecting approximately 10% female of reproductive age, yet\ndiagnosis is often delayed by 7 to 10 years due to the reliance on invasive laparoscopy as the reference standard\n[1]. According to World Health Organization (WHO), endometriosis cannot be prevented and there is no known\ncure; however, awareness and early diagnosis may be able to stop the natural course of the disease and lessen the\nlong-term effects of its symptoms. Still, diagnosing endometriosis and figuring out the factors affecting its course\nand related symptoms remains extremely difficult even now. The prolonged diagnostic process of endometriosis\nfrequently results in years of frustration and anxiety for sufferers. Treatment and care for these women are therefore\npostponed, sometimes for as long as 7 to 10 years after the onset of symptoms. Available evidence and data now\nsupport patients’ active participation in the detection and diagnosis of the illness. Relying on diagnosis through\npain itself is insufficient to identify endometriosis. Aside from the intricacy of the diagnosis and the prognosis\nhttps://doi.org/10.56578/ataiml050102\n11\n\nof symptoms, another concern is the percentage of women who show improvement following the surgery. The\nprognosis of reproductive issues in endometriosis-affected women is another urgent topic, thereby confirming that\nendometriosis has a significant impact on several aspects of women’s lives [1–4].\nNon-invasive methods, particularly those leveraging machine learning on imaging and clinical data, offer a\npromising pathway to earlier intervention and improved quality of life [4]. Recent advances have demonstrated the\npotential of artificial intelligence (AI) for endometriosis diagnosis in single-modality settings. Convolutional neural\nnetworks trained on Abdominal and Pelvic Magnetic Resonance Imaging (MRI) have shown strong lesion localization\nand diagnostic performance when paired with explainable AI (XAI) techniques such as Gradient-weighted Class\nActivation Mapping (Grad-CAM) for visual interpretability [5]. Similarly, dense neural networks applied to structured\nclinical data, augmented with SHapley Additive exPlanations (SHAP), have yielded competitive performance while\noffering feature-level transparency [6]. While effective in isolation, these features reflect structural manifestations,\nwhereas clinical data reflect patients’ history and symptomatology.\nIntegrating multimodal data addresses the limitations of single-modality systems by leveraging complementary\ninformation. Decision-level fusion, which combines output from independently trained models, offers a practical\nstrategy with reduced complexity compared with feature-level fusion. Despite its promise, multimodal fusion for\ndiagnosis of endometriosis has not yet been systematically evaluated with respect to calibration, robustness, and\nclinical benefit.\n2 Related Works\nNon-invasive diagnosis for endometriosis has greatly been improved via AI and Machine Learning (ML), offering\nup to 90% sensitivity and specificity in research settings [7]. Models have been developed using diverse data sources\nwhich include clinical features, self-reported symptoms [8], imaging (ultrasound or MRI), genetic, and proteomic\ndata.\nStolz et al. [9] found that a fused 3D T1/T2 MRI protocol achieved <94% accuracy in diagnosis and improved\ninter-reader reproductivity compared with standard 2D MRI sets. In the AI domain, a study applied AI-segmentation\n(fuzzy C-means clustering) MRI for ovarian endometriosis; this showed improved sensitivity and reduced processing\ntime. However, these efforts remain largely single-modality and rarely integrate structured clinical data or metrics\nfor decision analysis. Parallel to image-only approaches, models of structured clinical data have gained traction in\n2025. Fatade et al. [6] worked on explainable AI using SHAP-features for interpretation of clinical variables for\nendometriosis and demonstrated interpretable risk prediction. Yet models based purely on clinical features often\nunderperform in capturing spatial and anatomical cues available in MRI, and they suffer from calibration biases.\nThere is more evidence to support multimodal fusion methods that integrate sources of heterogeneous data\nfor improved diagnosis. Surveys of HER + imaging fusion underline consistent performance gains of multimodal\nsystems over single modalities. Mohsen et al. [10] similarly reviewed multimodal medical image fusion (MMF),\noutlined pixel-, feature-, and decision-level fusion strategies, and emphasized the transition toward deep learning-\nbased networks. Zubair et al. [11] explored specific image fusion studies such as MRI + Computed Tomography (CT)\nfusion using GANs and interval-gradient convolutional neural network (CNN) fusion methods further demonstrated\nthe benefits of combining anatomical and functional imaging. Yet in gynecologic AI research, application of\ndecision-level multimodal fusion remains scarce.\nCalibration and evaluation of clinical utility have emerged as essential topics in predictive modelling. The\nmethodology of decision curve analysis (DCA) has been recommended for assessing net clinical benefit of diagnostic\nmodels beyond discrimination alone [12]. Moreover, studies applying DCA to calibrated CNN illustrated that\nimprovements of calibration translate into meaningfully clinical benefit [13]. Although many endometriosis models\nreport discrimination (area under the curve (AUC), and accuracy), few incorporates calibration curves, expected\ncalibration error (ECE)/Brier metrics or DCA.\nCompared with pixel-level and feature-level fusion strategies, decision-level multimodal fusion offers several\nadvantages for clinical deployment. By combining independently optimized modality-specific models at the output\nstage, decision-level fusion reduces the need for extensive cross modal data harmonization and lowers computational\noverhead. Importantly, it preserves the interpretability of the model by allowing modality specific explanations for\nexample imaging heatmaps and clinical feature attributions to be presented separately, in order to align with clinical\ndecision-making workflows [14, 15].\nIn this study, we investigated decision-level fusion strategies that combined on an MRI-based U-Net model\nwith a structured clinical-data neural network. We comparatively evaluated weighted averaging, stacking, and\nconfidence-gating approaches to examine how different fusion strategies influenced diagnostic performance, proba-\nbility calibration, and clinical net benefit. Beyond conventional accuracy metrics, this work focused on elucidating\nthe trade-offs between predictive performance, robustness, interpretability, and implementation complexity, with the\naim of informing the selection of fusion strategies that are most suitable for real-world clinical decision-support\nsettings.\n12\n\n3 Methods\nThis study adopted both structured and unstructured data types which were obtained from patients with\nindications of endometriosis. These datasets included Abdominal and Pelvic MRI, symptoms, findings from\nphysical examination, and patients’ clinical history. In this study, we endeavored to address the gaps by comparing\ndecision-level fusion strategies (weighted averaging, stacking, and confidence-gating), assessing calibration via ECE\nBrier score, and quantifying net clinical benefit through DCA.\n3.1 Data and Preprocessing\nThis retrospective study included records for 1208 patients who underwent Abdominal and Pelvic MRI for\nsuspected endometriosis at Crestview Radiology Center across their branches in four cities in Nigeria (Lagos,\nKano, Ibadan, and Ilorin), where these reports were anonymized. Diagnostic labels were derived from a combi-\nnation of radiological reports and clinical documentation recorded in the electronic medical record. Specifically,\nendometriosis-positive cases were identified based on a consensus by gynecologists and radiologists following\nthe interpretation of MRI and clinical evaluation. While laparoscopic confirmation was considered the reference\nstandard, such data were not available in this cohort. The reliance on MRI-supported clinical diagnosis reflects real-\nworld diagnostic pathways and aligns with the current trend towards reducing invasive approach adopted in previous\nimaging-based AI studies for endometriosis and other gynecological disorders [16, 17]. Due to the retrospective\nnature of our dataset, there was no a priori calculation of the sample size. The size of the final cohort was determined\nby the availability of data during the study period.\n3.1.1 MRI data\nAbdominal-Pelvic examinations were acquired using 1.5T and 3T scanners under standardized protocols that\nincluded T2-weighted fast pin echo sequences. Lesions were segmented semi-automatically using 3D Slicer (open-\nsource platform, www.slicer.org), with manual refinements performed by experienced radiologists. The resulting\nlesion masks were used both to support the training of the model and to evaluate the localization performance of\nGrad-CAM heatmaps.\n3.1.2 Clinical data\nStructured clinical data included demographics (age, body mass index (BMI), and parity), gynecological and\nsurgical history, and profiles of symptoms (severity of pain, dysmenorrhea, and dyspareunia). Categorical features\nwere one-hot encoded, while continuous features were Z-scored normalized. Missing values were imputed using the\nmedian (continuous) or most frequent category.\n3.2 Base Predictive Models\n3.2.1 MRI model\nA dual-attention U-Net was employed for segmentation and classification of lesion. The architecture incorporated\nspatial and channel attention blocks to enhance the representation of features [5]. Training employed the Adam\nOptimizer with a learning rate of 1e-4, Dice loss for segmentation, and binary cross-entropy for classification. To\nenhance generalization, data augmentation included rotations, flips, and intensity scaling. For interpretability, Grad-\nCAM maps were generated from the classification head to highlight image regions most influential to predictions\n[5].\n3.2.2 Model of clinical data\nA densely feedforward neural network was trained using structured clinical features. Selection of features was\nguided by SHAP importance ranking to reduce dimensionality and mitigate overfitting [6]. The network comprised\nthree hidden layers (128-64-32 neurons) with ReLU activation and dropout regularization ( p = 0 .3). The Adam\nOptimizer (learning rate of 1e-3) was used with binary cross-entropy loss. Probabilistic output indicated the\nlikelihood of endometriosis.\n3.2.3 Model training and data splitting\nThe dataset was randomly partitioned into training (80%) and test (20%) sets at the patient level to prevent\nleakage of data. Randomization was performed using a fixed random seed to ensure reproducibility. The training\nset was used for model fitting and internal tuning, while the test set was held out and used exclusively for evaluation\nof final performance. Both MRI-based and clinical models were trained independently using the same data split, to\nensure fair comparison and consistent evaluation.\n13\n\n3.3 Fusion Strategies\nThree decision-level fusion strategies were evaluated:\na. Weighted Averaging (baseline): probabilities from both models were combined as:\nPfusion = αPMRI + (1 − α)Pclinical\nwhere, α was tuned on the validation set. The weighting parameter was selected on the validation set with a discrete\ngrid search over values between 0 and 1. The value yielding the best validation performance was fixed and applied\nto the test set to avoid leakage of information.\nb. Stacking Meta-Leaners: A logistic regression model was trained on the probabilistic output of the two base\nmodels. Five-fold cross-validation was applied to prevent overfitting.\nc. Confidence-Gating: For each case, the output of the model with higher prediction confidence (probability\nfarther from 0.5) was selected as the fused output.\n3.4 Evaluation Protocol\nTo thoroughly assess the performance of the proposed multimodal fusion framework, we evaluated models\nalong three complementary dimensions: predictive accuracy, probability calibration, and clinical utility. Predictive\naccuracy captures the ability of the models to correctly classify cases, while calibration reflects the reliability of\npredicted probabilities in reflecting the risk of the disease. Clinical utility was quantified using decision curve\nanalysis to determine whether fused models provided a measurable benefit in realistic diagnostic decision-making\nscenarios. Statistical testing was performed to determine whether observed improvements in fusion were significant\ncompared with single modality baselines.\nThe subsection that follows describes the metrics, calibration methods, decision curve framework, and statistical\nanalysis plan.\n3.4.1 Performance metrics\nThe performance metrics for the model was derived using accuracy, precision, recall, F1-score, and AUC.\n3.4.2 Calibration metrics\nThis was quantified using Brier score, ECE, and reliability diagram. To improve the probability further, two\npost-hoc recalibration methods were applied on the validation set. Platt scaling which fits regression model to map\nraw prediction scores into well-calibrated probabilities, making it effective when miscalibration follows a sigmoid\npattern [18]. Isotonic regression, however, is a non-parametric monotonic mapping that can flexibly correct irregular\ncalibration error [19, 20]. Using both approaches allowed us to compare a simple parametric correction with a more\nflexible, data-driven alternative.\n3.4.3 Clinical utility\nDCA was conducted to assess net clinical benefit across threshold probability from 0.2–0.6, a range relevant for\ngynecological triage decisions.\nThe probability range of 0.2–0.6 for the threshold was selected to reflect clinically plausible triage scenarios in the\nnon-invasive assessment of suspected endometriosis. Lower thresholds (around 20%) correspond to settings where\nclinicians prioritize sensitivity to avoid missed diagnoses, while higher thresholds (up to 60%) reflect scenarios\nrequiring greater diagnostic confidence before recommending further investigation or referral from specialists. This\nrange captures the decision space where risk-based stratification is most likely to influence clinical management.\n3.4.4 Statistical analysis\nBootstrapping with 1,000 iterations was used to compute 95% confidence intervals for performance metrics.\nPaired comparisons of accuracy were conducted using McNemar’s test, while AUCs were compared using the\nDeLong test. A p-value <0.05 was considered statistically significant.\n4 Results\n4.1 Overall Performance\nTable 1 summarizes the performance of the single modality and fusion models. The dual-attention U-Net trained\non MRI achieved an accuracy of 0.87, precision of 0.85, recall of 0.88, and F1-score of 0.86, while the dense neural\nnetwork trained on structured clinical data achieved an accuracy of 0.83, precision of 0.82, recall of 0.8, and F1-score\nof 0.81.\n14\n\nAll fusion strategies outperformed the single-modality baselines. Weighted average achieved an accuracy of 0.89,\nprecision of 0.89, and recall of 0.87. Stacking with logistic regression further improved the AUC to 0.92, representing\na statistically significant increase compared with single-modality model (DeLong test,p < 0.05). Confidence-gating\nmaintained competitive performance while prioritizing robustness under varying prediction of the confidence level.\nTable 1. Performance of single-modality and fusion models\nModel Accuracy Precision Recall F1-Score AUC\nMRI only (U-Net) 0.87 0.85 0.88 0.86 0.90\nClinical only (Dense NN) 0.83 0.82 0.80 0.81 0.86\nFusion-Weighted Average 0.89 0.89 0.87 0.86 0.91\nFusion-Stacking 0.90 0.90 0.88 0.89 0.92\nFusion-Confidence-Gating 0.88 0.87 0.86 0.86 0.90\nNote: Reported metrics represent point estimates evaluated on the held-out test set. Statistical uncertainty and stability of the result were\nassessed using paired hypothesis testing and resampling-based analysis, as described in the Methods section\nAlthough confidence intervals were not explicitly tabulated for all metrics, consistent performance improvements\nacross fusion strategies and statistically significant comparative tests support the robustness of the reported results.\n4.2 Calibration Performance\nIt is seen in Figure 1 (Reliability diagram) that both single models were modestly mis-calibrated, thus tending to\noverestimate probabilities at high confidence. The MRI model had an ECE of 0.08 and the clinical model was 0.1.\nFusion through stacking demonstrated the best calibration, with an ECE of 0.04 and the lowest Brier score of 0.09.\nWeighted averaging also improved calibration compared to single modality models, with an ECE of 0.05.\nThere is a distinct calibration pattern across the models from the reliability diagram. The MRI only model\nexhibited mild over-confidence in higher predicted probability bins, with predicted risks exceeding observed outcome\nfrequencies. In contrast, the clinical only model tended to underestimate risk in lower probability intervals, thus\nreflecting conservative probability output. The stacked fusion model demonstrated improved alignment between the\npredicted and observed probabilities across most bins, indicating more stable calibration behavior across the full\nprobability range.\nFigure 1. Reliability diagram of MRI (orange), clinical (blue), and fusion models (green). The stacking-based\nfusion model demonstrated the closest alignment between predicted and observed probabilities\n15\n\n4.3 Clinical Utility\nDecision curve analysis in Figure 2 shows that fusion models consistently achieved higher net benefit across\nclinically relevant probability thresholds (0.2–0.6). At a threshold of 0.4, stacking avoided approximately 12 to 18\nunnecessary invasive procedures per 100 patients, compared with the best-performing single modality model.\nFigure 2. Decision curve analysis comparing net clinical benefit of single-modality and fusion models across the\nthreshold probability\n4.4 Robustness Analysis\nExperiments of robustness indicated that the performance of single-modality models degraded under simulated\ndistribution shifts (Table 2). For example, under intensity-scaled MRI data, the accuracy of U-Net dropped from 0.87\nto 0.82. In contrast, the confidence-gating fusion approach maintained stable performance, with reduction of less\nthan 3% (from 0.88 to 0.86) in accuracy. Scanner-level cross-validation confirmed that fusion models generalized\nbetter across 1.5T and 3T subsets than single-modality baselines.\nTable 2. Results of robustness under distribution shifts and cross-scanner validation\nModel Accuracy Baseline Accuracy under Shift ∆ Accuracy\nMRI only 0.87 0.82 -0.05\nClinical only 0.83 0.78 -0.05\nFusion-Weighted 0.89 0.86 -0.03\nFusion-Stacking 0.90 0.87 -0.03\nFusion-Gating 0.88 0.86 -0.02\n4.5 Ablation Study\nThe result of the ablation study as shown in Table 3 shows the contribution of attention mechanisms and feature\nselection. When dual attention from U-Net was removed, there was a decrease in accuracy from 0.87 to 0.82.\nExcluding the selection of SHAP-based features could reduce the performance of the clinical model from 0.83 to\n0.79. Fusion models incorporating these components consistently achieved the strongest results.\nTable 3. Ablation study evaluating the impact of model components\nModel Variant Accuracy F1-Score AUC\nU-Net (No attention) 0.82 0.81 0.86\nDense NN (All features, no SHAP) 0.79 0.78 0.83\nFusion (Full model) 0.90 0.89 0.92\n16\n\n5 Discussion\nDemonstrated in this study is the fact decision-level multimodal fusion of MRI-based and clinical data-based\nmodels significantly improved the non-invasive diagnosis of endometriosis, when compared with single-modality\nsystems. Across the accuracy, calibration and clinical utility metrics, fusion strategies consistently outperformed\nindividual models. Stacking fusion achieved the best overall performance, with an AUC of 0.92, an expected\ncalibration error of 0.04, and the highest net clinical benefit in decision curve analysis.\nBeyond performance gains, model interpretability plays a critical role in supporting the decision making of\nclinicians. Grad-CAM visualizations from the MRI-based model highlight image regions contributing most strongly\nto predictions, thus enabling radiologists to verify whether model attention aligns with known anatomical sites\ninvolving endometriosis. This could increase confidence in the output of the model and assist in identifying subtle\nimaging patterns that may warrant closer inspection. Similarly, SHAP-based explanations from the clinical model\nprovide feature-level attribution, allowing clinicians to understand how profiles of symptoms, medical history, and\ndemographic factors contribute to estimates of individual risk. When used in conjunction with clinical judgement,\nthese interpretability output could support transparent risk stratification, facilitate clinician-AI collaboration, and\nimprove trust in AI-assisted non-invasive diagnostic workflows.\n5.1 Interpretation of Findings\nThe results highlight the complementary strengths of imaging and clinical data in characterizing endometriosis.\nMRI features capture structural manifestations of the disease, while clinical variables reflect symptoms burden and\nmedical history. Their integration yielded synergistic gains, particularly in calibration, hence suggesting that fused\nmodels generate more reliability estimates. This is clinically meaningful, as calibrated probabilities better align with\nreal-world decision thresholds. Importantly, decision curve analysis indicated that fusion models could prevent 12\nto 18 unnecessary invasive procedures per 100 patients, compared with MRI or clinical models alone. Thus, the\npotential of meaningful patient impact was underscored.\n5.2 Comparison with Prior Works\nPrevious studies largely focused on unimodal AI approaches for endometriosis, including convolutional networks\napplied to MRI [5], and dense neural network applied to structured data [6], to mention a few examples. While\nthese models achieved competitive performance in isolation, our findings showed that decision-level fusion offered\nincremental benefit. These results align with broader multimodal AI literature, where fusion has improved the\nreliability of diagnosis of oncology, cardiology, and neuroimaging tasks [21, 22].\n5.3 Clinical Implications\nThe improved calibration of the fusion models is especially relevant in gynecological practice, where decisions of\ntreatment often hinge on threshold probability. A well-calibrated system reduces the risk of over- or under-treatment,\nand it enables clinicians to use model output as reliable adjuncts in triage and surgical planning.\n5.4 Limitations\nThis study has several limitations. First, diagnostic labels were derived from MRI reports and clinical consensus\nrather than laparoscopic or histopathological confirmation. While this reflects real-world diagnostic practices,\nit may introduce label noise. Second, MRI segmentation was performed using semi-automatic approach with\nexperts’ manual correction, but formal inter-observers’ agreement metrics (e.g., Dice similarity coefficient) were\nnot evaluated. While expert-guided segmentation reflects realistic workflows of clinical annotation, future studies\ncould incorporate independent multi-rater annotations and quantitative agreement analysis to further strengthen the\nreliability of segmentation. Third, data were obtained from a single institution, which may limit generalizability.\nFourth, while we explored decision-level fusion methods, feature-level or representation-level fusion strategies may\noffer further gains and should be investigated. Fifth, although missingness was limited, simple imputation strategies\nmay attenuate associations for certain features, and future work could explore more advanced imputation methods.\nFinally, it is important to mention the likely imitation that comes with single diagnostic center. This may limit the\ngeneralizability of the findings to other institutions with differing patient populations, imaging protocols, and clinical\nworkflows. Variations in scanner vendors, acquisition parameters, and practices of documentation when deployed,\ncould affect the performance of the model. However, the employment of decision-level fusion partially mitigates\nthese challenges by allowing independent optimization of modality-specific models and reducing dependencies on\ntightly coupled cross-model feature representations. Future work will focus on external validation across multiple\ncenters, including settings with heterogeneous imaging infrastructure, availability of varying levels of clinical data\nas well as prospective evaluation to assess robustness under conditions of real-world deployment.\n17\n\n5.5 Future Work\nFuture work should pursue prospective and multi-center validation of multimodal fusion models, incorporating\nboth clinical and imaging data streams. Additional research is needed to explore dynamic weighing strategies,\nintegrate temporal clinical data, and evaluate feature-level fusion approaches. Importantly, embedding these models\nin clinical-facing decision support systems and assessing their applicability, trust, and interpretability in real-world\nworkflows would be critical for translation into practice.\n6 Conclusions\nDecision-level multimodal fusion of MRI-based and clinical data-based AI models enhance the non-invasive\ndiagnosis of endometriosis by improving diagnostic accuracy, probability calibration, and clinical utility, compared\nwith single-modality approaches. Among the evaluated strategies, stacking-based fusion achieved the strongest\noverall performance and calibration, indicating high potentially clinical value. These findings support the integration\nof multimodal AI into gynecological decision support systems and highlight the demand for prospective and multi-\ncenter validation to enable translation into clinical practice.\nDespite superior performance, stacking-based fusion entails greater complexity in implementation, including\ncoordinated integration of the model and adaptation of the workflow. Simpler fusion strategies, such as weighted\naveraging and confidence-gating, may offer advantages in case of deployment, computational efficiency, and clinical\nacceptance, particularly in resource-constrained settings while maintaining competitive performance. Collectively,\nthese findings underscore the importance of balancing diagnostic accuracy with clinical reliability, interpretability,\nand feasibility when translating multimodal AI systems into routine practice.\nAuthor Contributions\nConceptualization, study design, software implementation, formal analysis, data curation, visualization, liter-\nature review, and writing—original draft preparation, O.B.F.; radiology analysis, MRI lesion annotation, manual\nsegmentation review, and provision of clinical expertise, O.F.A.; methodology development and preparation of\nthe methodology section, F.A.S. and K.I.O.; research documentation, data organization, and compilation of study\nmaterials, K.I.O. and G.C.O.; manuscript preparation, section integration, and overall structuring of content, G.O.O.\nand O.B.F.; abstract writing, manuscript refinement, and intellectual review support, O.B.F. and F.A.S. All authors\nhave read and agreed to the published version of the manuscript.\nData Availability\nThe data used to support the research findings are available from the corresponding author upon request.\nConflicts of Interest\nThe authors declare that they have no conflicts of interest.\nReferences\n[1] World Health Organization, “Endometriosis,” 2025. https://www.who.int/news-room/fact-sheets/detail/endom\netriosis\n[2] F. Gkrozou, O. Tsonis, F. Sorrentino, L. Nappi, A. Vatopoulou, C. Skentou, S. Pandey, M. Paschopoulos, and\nA. Daniilidis, “Endometriosis predictive models based on self-assessment questionnaire, clinical examination,\nor imaging findings: A narrative review,”J. Clin. Med., vol. 13, no. 2, p. 356, 2024. https://doi.org/10.3390/jc\nm13020356\n[3] M. Szubert, A. Rycerz, and J. R. Wilczy ´nski, “How to improve non-invasive diagnosis of endometriosis with\nadvanced statistical methods,”Medicina, vol. 59, no. 3, p. 499, 2023. https://doi.org/10.3390/medicina590304\n99\n[4] Q. J. Hudson, A. Perricos, R. Wenzl, and I. Yotova, “Challenges in uncovering non-invasive biomarkers of\nendometriosis,” Exp. Biol. Med., vol. 245, no. 5, pp. 437–447, 2020. https://doi.org/10.1177/15353702209032\n70\n[5] A. O. Kuyoro, O. B. Fatade, and E. E. Onuiri, “Enhancing non-invasive diagnosis of endometriosis through\nexplainable artificial intelligence: A Grad-CAM approach,”Acadlore Trans. AI Mach. Learn., vol. 4, no. 2, pp.\n97–108, 2025. https://doi.org/10.56578/ataiml040203\n[6] O. B. Fatade, A. O. Kuyoro, and E. E. Onuiri, “Explainable AI for endometriosis diagnosis: A dense neural\nnetwork approach with SHAP interpretation,” Int. J. Res. Stud. Inf. Technol., 2025. https://doi.org/10.51244/I\nJRSI.2025.12030070\n18\n\n[7] S. Bendifallah, A. Puchar, S. Suisse, L. Delbos, M. Poilblanc, P. Descamps, F. Golfier, C. Touboul, Y. Dabi,\nand E. Darai, “Machine learning algorithms as a screening approach for patients with endometriosis,”Sci. Rep.,\nvol. 12, no. 1, p. 639, 2022. https://doi.org/10.1038/s41598-021-04637-2\n[8] A. Goldstein and S. Cohen, “Self-report symptom-based endometriosis prediction using machine learning,”\nSci. Rep., vol. 13, no. 1, p. 5499, 2023. https://doi.org/10.1038/s41598-023-32761-8\n[9] A. Stolz, L. F. Pupulim, M. Rojas Soldado, P. Chabloz, and K. Kinkel, “Fusion 3D T1/T2 MRI for diagnosing\npelvic deep infiltrating endometriosis: A non-inferiority study,” Eur. J. Radiol., vol. 187, p. 112091, 2025.\nhttps://doi.org/10.1016/j.ejrad.2025.112091\n[10] F. Mohsen, H. Ali, N. El Hajj, and Z. Shah, “Artificial intelligence-based methods for fusion of electronic health\nrecords and imaging data,” Sci. Rep., vol. 12, no. 1, p. 17981, 2022. https://doi.org/10.1038/s41598-022-225\n14-4\n[11] M. Zubair, M. Hussain, M. A. Al-Bashrawi, M. Bendechache, and M. Owais, “A comprehensive review of\ntechniques, algorithms, advancements, challenges, and clinical applications of multi-modal medical image\nfusion for improved diagnosis,”arXiv, vol. 2505, p. 14715, 2025. https://doi.org/10.48550/arXiv.2505.14715\n[12] A. J. Vickers and F. Holland, “Decision curve analysis to evaluate the clinical benefit of prediction models,”\nSpine J., vol. 21, no. 10, pp. 1643–1648, 2021. https://doi.org/10.1016/j.spinee.2021.02.024\n[13] D. Deniffel, N. Abraham, K. Namdar, X. Dong, E. Salinas, L. Milot, F. Khalvati, and M. A. Haider, “Using\ndecision curve analysis to benchmark MRI-based deep learning models for prostate cancer risk assessment,”\nEur. Radiol., vol. 30, no. 12, pp. 6867–6876, 2020. https://doi.org/10.1007/s00330-020-07030-1\n[14] C. Cui, H. C. Yang, Y. H. Wang, S. L. Zhao, Z. Asad, L. A. Coburn, K. T. Wilson, B. A. Landman, and Y. Huo,\n“Deep multimodal fusion of image and non-image data in disease diagnosis and prognosis: A review,” Prog.\nBiomed. Eng., vol. 5, no. 2, p. 022001, 2023. https://doi.org/10.1088/2516-1091/acc2fe\n[15] S. C. Huang, A. Pareek, S. Seyyedi, I. Banerjee, and M. P. Lungren, “Fusion of medical imaging and electronic\nhealth records using deep learning: A systematic review and implementation guidelines,” NPJ Digit. Med.,\nvol. 3, no. 1, p. 136, 2020. https://doi.org/10.1038/s41746-020-00341-z\n[16] A. Kido, Y. Himoto, Y. Moribata, Y. Kurata, and Y. Nakamoto, “MRI in the diagnosis of endometriosis and\nrelated diseases,” Korean J. Radiol., vol. 23, no. 4, p. 426, 2022. https://doi.org/10.3348/kjr.2021.0405\n[17] M. Bazot and E. Darai, “Diagnosis of deep endometriosis: Clinical examination, ultrasonography, magnetic\nresonance imaging, and other techniques,”Fertil. Steril., vol. 108, no. 6, pp. 886–894, 2017. https://doi.org/10\n.1016/j.fertnstert.2017.10.026\n[18] C. Gupta and A. Ramdas, “Online Platt scaling with recalibration,” in Proceedings of the 40th International\nConference on Machine Learning, Honolulu, Hawaii, USA, 2023, pp. 12 182–12 204. https://proceedings.mlr.\npress/v202/gupta23c.html\n[19] E. Berta, F. Bach, and M. I. Jordan, “Classifier calibration with ROC-Regularized Isotonic Regression,” in\n27th International Conference on Artificial Intelligence and Statistics, Valence, Spain, 2024, pp. 1972–1980.\nhttps://hal.science/hal-04295601\n[20] L. Huang, J. Zhao, B. Zhu, H. Chen, and S. Van den Broucke, “An experimental investigation of calibration\ntechniques for imbalanced data,”IEEE Access, vol. 8, pp. 127 343–127 352, 2020. https://doi.org/10.1109/AC\nCESS.2020.3008150\n[21] Y. D. Zhang, Z. C. Dong, S. H. Wang, X. Yu, X. J. Yao, Q. H. Zhou, H. Hu, M. Li, C. Jim´enez-Mesa, J. Ramirez,\nand et.al, “Advances in multimodal data fusion in neuroimaging: Overview, challenges, and novel orientations,”\nInf. Fus., vol. 64, pp. 149–187, 2020. https://doi.org/10.1016/j.inffus.2020.07.006\n[22] J. Lipkova, R. J. Chen, B. Chen, M. Y. Lu, M. Barbieri, D. Shao, A. J. Vaidya, C. Chen, L. Zhuang,\nD. Williamson, and et.al, “Artificial intelligence for multimodal data integration in oncology,” Cancer Cell,\nvol. 40, no. 10, pp. 1095–1110, 2022. https://doi.org/10.1016/j.ccell.2022.09.012\n19","source_license":"CC0","license_restricted":false}