Diagnostic Performance of Computed Tomography-Based Machine Learning Models in the Classification of Adnexal Masses - A Systematic Review

Introduction

Adnexal masses refer to tumoral formations that originate from the ovaries, fallopian tubes, or the surrounding structures like para-ovarian cysts and polyps, found in females of all ages but especially in the reproductive ages. Also, the masses can originate from functional or non-functional tumors due to physiological changes and inflammatory conditions of benign and malignant neoplasms.¹ The potential malignancy underscores the importance of early, precise, and prompt diagnosis to reduce associated morbidity and mortality.² “Adnexal masses” are commonly found during imaging studies of the pelvis. In some instances, particularly those that are not as common, a mass might present with acute or intermittent pain. In the general population, the prevalence of “adnexal masses” cannot be known since most of the adnexal masses remain asymptomatic and undiagnosed.¹

Currently, the difference between benign and malignant “Adnexal masses” is primarily determined by their imaging characteristics.^3–5 Ultrasound (USG) is often used as an imaging modality for the evaluation and characterization of adnexal masses based on non-invasive properties and accessibility. It has limitations in terms of dependency and resolution of inconsistency, affecting its sensitivity for distinguishing between benign and malignant masses.⁶ CT is frequently utilized in routine clinical practice for the incidental initial detection of conditions due to its spatial resolution, broad accessibility, and shorter acquisition duration.⁷ The characterization of adnexal masses has traditionally depended upon these imaging modalities: techniques and subjective assessments. Nevertheless, there are limitations in evaluating the heterogeneity of masses. Thus, it is important to use a precise, objective, non-invasive approach for the categorization of adnexal masses using CT imaging as it offers higher sensitivity compared to USG, performs nearly at par with MRI, and provides the additional advantage of rapid acquisition.⁸

The algorithms used by artificial intelligence (AI) have the ability to scrutinize complex image information and enable the early identification and characterization of lesions using image recognition and the detection of minute details that may not be observable by the human eye.^9–13 However, CT-based AI models have excellent accuracy and specificity in classifying lesion, which helps in cancer imaging and treatment monitoring.^14–16

This review aims to enhance the detection, classification, and characterization of adnexal masses, thereby assisting radiologists in providing more accurate diagnosis. These results may help the gynecologists to choose more appropriate and personalized therapeutic approaches that could improve clinical outcomes and reduce disease aggressiveness in patients with adnexal masses.

Methods

Results

Selections approaches

After duplicate and abstract removal, a total of 1107 original studies were retrieved, and 12 were found to be eligible for full-text screening. Of these, articles are part of the review that fulfilled the inclusion criteria. This process of selection is elaborated in Figure 1.

Figure 1. PRISMA flow chart for the articles included in the review.

Risk of bias analysis

The quality of the studies included was assessed using the QUADAS-2. Generally, there was a low risk of bias in the domains of patient selection, index test, and reference standard, which is an indication of high methodological quality showed in Figure 2.

Figure 2. QUADAS-2 analysis.

All the studies included in the review, namely Yu et al.,²¹ Li et al.,²⁵ Jan et al.,²⁸ Li et al.,²³ Linton-Reid et al.,²⁰ Leng et al.,²⁷ Chen et al.,²⁴ Yu et al.,²⁹ Su et al.,³⁰ and Liu et al.,²⁶ had a low risk of bias in the domain of patient selection, which is an indication that the studies had appropriate study populations and that there was no selection bias. The domains of index test and reference standard also had low risks of bias, which is an indication that the studies applied the tests appropriately and that they used accepted diagnostic reference standards. Low risk was found in most studies in the flow and timing domain, reflecting appropriate intervals between the index test and reference standard. However, a moderate risk was found in this domain by Leng et al.,²⁷ which could be attributed to differences in follow-up or reporting.

However, concerns regarding applicability were found to be high in most studies, primarily because of the single-center study nature, lack of heterogeneity in the population, and differences in imaging protocols and model validation approaches. Only Li et al.²³ and Leng et al.²⁷ reported a moderate level of concerns regarding applicability. Thus, although the internal validity was excellent, external validity is poor, and there is a need for multicenter, externally validated studies.

Discussion

This systematic review draws attention to the increasing importance of CT radiomics and machine learning models in the evaluation of adnexal masses, especially in differentiating benign from malignant ovarian tumors. In general, most of the models used in the studies had moderate to excellent performance, which indicates that image analysis can provide important information beyond visual inspection.

The vast majority of the included studies used contrast-enhanced CT scans, and there was a strong preference for the portal venous phase, as reported by Yu et al.²¹ and Li et al.³¹ The portal venous phase offers more stable lesion enhancement and the ability to visualize tumor heterogeneity, which is essential for radiomics analysis. The studies using this phase reported significantly higher AUC values, as reported in the earlier imaging literature that suggests portal venous CT as the optimal phase for ovarian tumor assessment.

With respect to analytical methods, ensemble or hybrid methods tended to perform better than single algorithm classifiers. Li et al.,²⁵ reported that the use of multiple classifiers in ML (random forest, support vector machine, and multi-layer perceptron) resulted in an AUC of 0.96, which was superior to the performance of individual classifiers. Likewise, Li et al.²³ showed that the use of nomogram-based methods, which integrated radiomics and logistic regression, was superior in terms of robustness, with a balance between high accuracy and interpretability. By contrast, single-method classifiers like the radiomics-SVM model used by Yu et al.²¹ tended to perform relatively poorly (AUC 0.86), suggesting a lack of ability to model the complexity of tumors.

The models that integrated deep learning (DL) performed very well in more complex clinical tasks. Liu et al.²⁶ combined the radiomics approach with a ResNet-18 architecture to predict peritoneal metastasis with an AUC of 0.95 and high sensitivity and specificity. Jan et al.²⁸ also combined the 3D U-Net-derived features, showing that the DL approach can extract spatial and hierarchical tumor information that may not be captured by handcrafted radiomics alone. These results are in line with Park et al.,³² who found that the combination of CT texture analysis with ML improved the detection of ovarian malignancy compared to radiologist assessment alone.

When contrasted with previous radiomics analysis reviews in the context of ovarian cancer imaging, the results of this review are consistent with the general consensus that tree-based and boosting methods (random forest, XGBoost, LightGBM) generally perform better than simpler distance-based approaches like K-nearest neighbors and naive Bayes. Previous studies that are not included in this review have also highlighted that hybrid clinicoradiomic models generally offer improved diagnostic performance compared to radiomics models alone.

However, some limitations were apparent despite the encouraging results. The majority of the studies were retrospective and single-center, which may pose a risk of selection bias and lack of generalizability. There was heterogeneity in the parameters of CT image acquisition, segmentation approaches (2D vs. 3D), feature selection algorithms, and validation procedures, making it difficult to compare the results and perform meta-analysis. Moreover, some of the models were not externally validated prospectively, which is essential for clinical use.

Future studies should focus on large-scale, prospective, multi-institutional studies with standardized CT acquisition and radiomics pipelines. The use of fully automated segmentation and end-to-end deep learning models may improve clinical applicability. External validation on different populations and scanner platforms is necessary before clinical application. The combination of radiomics analysis with clinical and genomic information may also help in individualized risk assessment and management of adnexal masses.

Conclusion

CT radiomics and machine learning algorithms have shown great potential as ancillary tools for the assessment of adnexal masses. The algorithms have shown high accuracy and could potentially help radiologists in distinguishing between benign and malignant masses, thus helping in appropriate management. However, before their widespread use, there is a need for further prospective studies. Once validated, these tools could help in improving the accuracy of diagnosis and thus help in personalized management in gynaecologic oncology.

Ethics and consent

This is a review article. Ethical approval and consent were not required.