Artificial Intelligence in Obstetric and Gynecological MR Imaging

Abstract

This review explores the significant progress and applications of artificial intelligence (AI) in obstetrics and gynecological MRI, charting its development from foundational algorithmic techniques to deep learning strategies and advanced radiomics. This review features research published over the last few years that has used AI with MRI to identify specific conditions such as uterine leiomyosarcoma, endometrial cancer, cervical cancer, ovarian tumors, and placenta accreta. In addition, it covers studies on the application of AI for segmentation and quality improvement in obstetrics and gynecology MRI. The review also outlines the existing challenges and envisions future directions for AI research in this domain. The growing accessibility of extensive datasets across various institutions and the application of multiparametric MRI are significantly enhancing the accuracy and adaptability of AI. This progress has the potential to enable more accurate and efficient diagnosis, offering opportunities for personalized medicine in the field of obstetrics and gynecology.

Introduction

Artificial intelligence and its history

The concept of artificial intelligence (AI), initially introduced by Alan Turing in 1950, has evolved significantly over recent decades. Turing’s “Turing test” proposed a method to assess a machine’s capability to exhibit human-like intelligence. In 1956, John McCarthy defined AI as the creation of intelligent machines. At its broadest level, AI encompasses computer algorithms designed to perform tasks that typically require human intelligence. This includes a wide range of capabilities, from understanding natural languages to recognizing patterns in complex data.¹ Early AI developments were based on simple “if, then” rules, but they gradually progressed to more complex algorithms, giving rise to subfields such as machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs).¹ These relationships and the main methods included within each are presented in Fig. 1.

Fig. 1.

The relationship between artificial intelligence, machine learning, deep learning, and convolutional neural networks, including the main methods within each. GNN, graph neural network; LASSO, least absolute shrinkage and selection operator; LSTM, long short-term memory network; PCA, principal component analysis; RNN, recurrent neural network; SVM, support vector machine.

Machine learning

ML is a part of AI that learns from data and improves with more exposure. This includes a wide range of capabilities, from understanding natural languages to recognizing patterns in complex data.¹ This learning process eliminates the need for explicit programming and allows the system to adapt to new and unforeseen scenarios. ML can be further divided into categories such as supervised learning, where the algorithm learns from a labeled dataset, and unsupervised learning, where the algorithm identifies patterns in unlabeled data.^1–3 Common ML methods include support vector machine (SVM) for classification, often used in tumor and disease diagnosis. Decision trees help in easy-to-interpret classification by splitting data based on features. Principal component analysis (PCA) reduces dimensionality for visualization and noise reduction. Least absolute shrinkage and selection operator (LASSO) regression improves accuracy by selecting features and regularizing models. Random forest is an ensemble method that combines multiple decision trees to improve accuracy and robustness by averaging their predictions.⁴

Deep learning

DL, a more specialized branch of ML, employs artificial neural networks with multiple layers, or “deep” networks, to process data in complex ways. DL models can automatically learn high-level features from data, which makes them particularly effective for tasks such as image recognition. These models mimic the way human brains operate, albeit in a simplified form, by processing information through a series of hierarchical layers, with each layer focusing on different aspects of the data.^1–3

Convolutional neural networks

CNN is a specific type of DL model especially suited for analyzing visual imagery. These networks are composed of convolutional and pooling layers structured to capture and abstract various levels of visual features. In the convolutional layers, filters are applied across the input data to identify basic features such as edges and textures. Following this, the pooling layers reduce the spatial size of the representation, enhance the network’s efficiency, and help extract more robust features. This structure enables to interpret and analyze images with high efficiency and accuracy.^1–3 One challenge with CNN is their need for large labeled datasets. Training involves backpropagation, which can cause overfitting. To counter this, techniques like transfer learning and data augmentation are used.^1–3 Transformer networks, known for their transfer capability, excel in image segmentation and classification. Combining CNN and Transformers has shown promising results in radiology.^5,6

Radiomics

Radiomics is an emerging field in medical imaging that extracts a large number of features from medical images using data characterization algorithms. These features, which can include various statistical, shape, and texture metrics, provide comprehensive quantification of tumor phenotypes. The integration of radiomics and AI, particularly through ML and DL techniques, has shown significant potential in enhancing medical imaging.³ ML methods such as PCA and LASSO are used for feature selection, SVM and random forest for model building, and evaluation metrics like cross-validation and receiver operating characteristic curves for performance assessment. DL leverages image preprocessing techniques like noise reduction and segmentation, automates feature extraction, and uses neural networks for model building. Specifically, CNNs excel in image preprocessing, automatic feature extraction, and handling specific image recognition tasks. Several recent publications have reported that radiomics and DL combination models demonstrate excellent predictive results, and it may be better for radiomics to cross-talk with DL as the integrated model.^7,8

Classification

Classification is the process of identifying which category or class a new observation belongs to based on a training set of data containing observations whose category is known. This is achieved by developing a model that makes predictions based on data features. Although radiomics is at the core of AI research on image classification, there has been a recent increase in DL studies that use the images themselves. An overview of radiomics and DL model development are shown in Figs. 2 and 3, respectively. Radiomics and DL model development includes training, validation, and testing phases. During training, the model learns from a labeled dataset to recognize patterns. The validation phase is used to tune the model settings and prevent overfitting by employing a separate dataset to evaluate the performance and optimize the parameters. Finally, the testing phase assesses the model’s effectiveness on a new dataset, the test set, providing an unbiased evaluation of its real-world applicability and ability to generalize to unseen data. This testing is sometimes referred to as validation, independent validation, or external validation; however, to avoid confusion, this review will refer to the use of data from other facilities as external validation.

Fig. 2.

Overview of the radiomics model development. VOI, volume of interest.

Fig. 3.

Overview of the deep learning model development.

Uterus

Uterine leiomyosarcoma

Uterine leiomyosarcoma (LMS) is highly malignant with a poor prognosis, and the overall 5 year survival rate for all stages ranges from 15% to 25%.⁹ A major challenge in diagnosing LMS is distinguishing it from atypical uterine leiomyomas. Although MRI is often used to establish a proper diagnosis, definitive differentiation remains difficult. Therefore, various AI studies are being conducted to differentiate between them;^10–17 however, owing to the nature of the disease, high diagnostic sensitivity is required for clinical applications.

For studies using ML-based analysis of quantitative imaging features from multiparametric MRI, in 2019, Nakagawa et al. reported that an ML analysis of 30 uterine sarcomas was more accurate than the interpretations of experienced radiologists, achieving an area under the curve (AUC) of 0.93 compared to 0.80.¹⁰ Furthermore, they found that another ML-based model using a dataset of 11 uterine sarcomas was superior to that of positron emission tomography (PET) imaging for diagnosis, with an AUC of 0.92 versus 0.85, however, these studies did not provide sensitivity metrics nor evaluation with an independent validation.¹¹ In 2020, Malek et al. reported that an ML algorithm exhibited 100% sensitivity, specificity, and accuracy in distinguishing between LMS and myomas based on 21 malignant cases. However, this method also lacked independent validation.¹²

For studies of ML models utilizing radiomics features, in 2019, Xie et al. demonstrated that an ML model utilizing radiomics features extracted from the apparent diffusion coefficient (ADC) of the entire uterus outperformed models based solely on features extracted from tumors or from tumors and a small piece of surrounding tissue in distinguishing between uterine sarcoma and myomas. The optimal radiomics model using the data from 29 uterine sarcomas achieved an AUC of 0.83; however, it had a sensitivity of only 76%. In addition, no independent validation was evaluated.^13,14 In 2021, Wang et al. demonstrated that their integrated ML model, which combined T2-weighted image (T2WI)-based radiomics with clinical data from 40 malignant uterine mesenchymal tumors, outperformed models using only T2WI-based radiomics in diagnosing the malignancy of uterine mesenchymal tumors, achieving an AUC of 0.91. This combined model also matched or exceeded the performance of 2 expert radiologists, although sensitivity metrics were not provided.¹⁵ In 2022, Dai et al. investigated the differentiation of uterine sarcomas from atypical leiomyomas using radiomic features derived from T2WI and diffusion-weighted imaging (DWI), along with clinical data from 86 sarcoma cases. The complex multiparameter models, which combined with age, menopausal status, abnormal vaginal bleeding, and mean ADC value, based on transfer learning, outperformed models based solely on radiomics. These models achieved a peak AUC of 0.96 and a sensitivity of 0.92.¹⁶

In a DL study using imaging itself in 2024, Toyohara et al. developed an automated diagnostic system that leveraged CNNs (algorithm: MobileNet-V2 network) trained by 55 cases of uterine sarcoma. This system attained an accuracy of 92% and sensitivity of 92%, utilizing unsegmented JPEG images and incorporating all multiparametric image slices that included the tumor.¹⁷

Despite all reports showing relatively high AUCs, the sensitivity remains insufficient. Given that LMS is a rare tumor, the limited number of LMS cases poses a challenge, necessitating future multicenter studies. Moreover, the inclusion of carcinosarcomas within sarcomas in many studies^{10,11,13,14,16} is thought to have implications for clinical application.

Endometrial cancer

Endometrial cancer (EC) accounts for the vast majority (>90% in the USA) of corpus uteri malignancies.¹⁸ Most AI developments have been in radiomics studies. Because of the numerous studies on radiomics in EC, this review focuses only on papers published after 2020. These evaluations have included differential diagnosis,^19–23 risk factors such as tumor grading,^24–27 myometrial invasion,^28–31 lymph node (LN) metastasis prediction,^32,33 other risk factors prediction,^34–38 combinations of these risk factors,^39–42 and outcome prediction.^43,44 While some studies have relied on a single imaging sequence, with T2WI being the most prevalent, the majority of recent studies have utilized multiparametric MRI. Additionally, multicenter studies with external validations have been increasing in recent years.^{22,23,27,35,38,39,43}

In multicenter studies that performed external validation using ML-based models utilizing radiomics data from multiparametric MRI, to evaluate the differentiation of tumor grade, in 2020, Yan et al. developed an ML-based nomogram that incorporated clinical information and radiomics data to preoperatively identify high-risk ECs. This study included 394 patient datasets for training and compared the treatment recommendations derived from the nomogram with the actual clinical decisions made for the patients. The results showed that radiologists assisted by the radiomics nomogram outperformed those relying solely on traditional methods. Moreover, the study reported an improvement in surgical planning with the radiomics nomogram, indicating that 11–15 of every 100 patients benefitted from better-tailored surgical interventions compared to the actual procedures performed.²⁷ To predict the recurrence risk, in 2023, Lin et al. developed a fusion ML model based on clinicopathological factors and radiomic features. In total, 235 patients were included in the training cohort, and fusion models showed better performance (AUC: 0.85) than models based on clinicopathological features (AUC: 0.75) or radiomics features alone (AUC: 0.78) in an external validation.⁴³

Furthermore, one of the strengths of radiomics is its ability to predict risk factors that cannot be identified with conventional imaging methods. In 2023, Song et al. assessed multiparametric MRI-based radiomics ML models for determining microsatellite instability (MSI) status in EC with 160 patients from single center in the training cohort. Among the several algorithms, SVM emerged as the most effective, achieving an AUC of 0.94 in an external validation.³⁴ In addition, in 2022, Lefebvre et al. discovered that a 3D radiomics ML model, which utilized multiparametric MRI from 157 patients at a single center, was effective in distinguishing various critical factors in EC. Specifically, the model differentiated deep myometrial invasion, lymphovascular space invasion (LVSI), tumor grade, and stage, achieving AUC values of 0.81, 0.80, 0.74, 0.74, and 0.84, respectively, for external validation.³⁹

The integration of radiomics and genomics, known as radiogenomics, investigates the correlation between imaging features and genomic data, including gene expression analysis. Radiogenomics holds great potential for predicting molecular features of tumors through easily accessible and cost-effective medical imaging. In 2023, Li et al. developed an ML model to predict MSI status and programmed death ligand-1 (PD-L1) expression using radiomics features from multiparametric MRI, obtained from both intratumoral and peritumoral regions, combined with clinicopathologic characteristics. Sixty patients from single center were used for model construction. The combination model achieved an AUC of 0.85 for predicting MSI status and 0.76 for predicting PD-L1 expression in the external validation.³⁸ In 2023, Lin et al. developed an ML-based radiomics model to preoperatively predict the pathogenic POLE mutation status, which helps tailor the surgical procedure and adjuvant treatment strategy. This single center study involved 35 POLE-mutant cases and 103 non-POLE-mutant cases. The radiomics model, which integrated features from multiparametric MRI, demonstrated high performance in the validation cohorts, with an AUC of 0.81.⁴⁵

Studies of image classification by DL using the images themselves have also been performed. In 2020, Chen et al. conducted a study to assess myometrial invasion and evaluate the diagnostic accuracy of a DL model. The model (algorithm: YOLOv3), trained on data from 313 patients with automated lesion detection and cropping, showed a higher accuracy (85%) than radiologists (78%). Notably, the highest accuracy (86%) was achieved when the model assessments were combined with those of the radiologists, although this study did not specify the type of imaging used.⁴⁶ In 2022, Urushibara et al. developed a CNN (algorithm: Xception) model to diagnose EC using data from 184 patients. The model processed various sequences of semi-automatically cropped JPEG images and demonstrated AUC values ranging from 0.88 to 0.95 for single image sets and 0.87 to 0.93 for combined sets, indicating diagnostic performance on par with or superior to that of radiologists.⁴⁷

EC, with its large number of cases, easily obtainable pathological results, and uniform and localized tumors, has undergone more advanced AI research, including multicenter studies, 3D analysis, and external validation, than other gynecologic tumors.

Cervical cancer

Cervical cancer (CC) is the fourth most common cancer worldwide and is one of the top three cancers in women aged <45 years in 145 countries.⁴⁸ Most studies have focused on risk prediction and outcome prediction. In radiomics research, due to the numerous studies on radiomics, this review focuses only on papers published after 2020, and studies predicting LN metastasis are the most common,^49–57 and several studies have predicted LVSI.^55,58–61 In research on locally advanced cancer, MRI-based^62–74 and combined PET-CT-based radiomics have primarily been used⁷⁵ with most of these studies reporting survival^{63,64,66–71,73,76} or response to therapy^62,71,72,77 as the main predictive outcomes of interest. In recent years, there has been an increasing number of multicenter studies that include external validation.^{51,56,61,75,76}

In multicenter studies with external validation that use ML or DL-based models utilizing radiomics data from multiparametric MRI, for predicting para-aortic LN involvement, in 2023, Lucia et al. developed ML models in patients with locally advanced CC before chemoradiotherapy using PET-CT and MRI radiomics features combined with clinical parameters. Data from 102 patients were used for training. The ComBat-radiomic model performed well in the external validations, with an AUC of 0.90.⁷⁵ For predicting risk factors that cannot be identified with conventional imaging methods, in 2023, Wu et al. developed and validated a nomogram based on MRI radiomics and the clinical signature to assess LVSI status. A total of 129 cases were used for training, and the nomogram showed high predictive performance in the external validation (AUC: 0.83) for predicting LVSI.⁶¹ For predicting recurrence risk factors in early-stage CC, in 2023, Zhang et al established and validated a DL radiomics nomogram based on intratumoral and peritumoral regions of MRI and clinical characteristics. Integrated with independent clinical factors, a nomogram was constructed using the data from 126 patients in the training cohort. This DL-based nomogram demonstrated satisfactory performance in predicting recurrence risk factors, with an AUC of 0.86 in the external validation.⁷⁶

For image classification by DL using the images themselves, in 2021, Urushibara et al. developed a CNN model (algorithm: Xception), which was trained using JPEG images of sagittal T2WI from 117 patients to diagnose CC. The model achieved an AUC of 0.93, demonstrating a diagnostic performance comparable to that of experienced radiologists.⁷⁸

All models in the published studies have achieved high AUCs; however, there is a clinical application challenge in that there are no studies covering all stages of CC, as treatment methods vary by stage.

Ovary

In 2018, ovarian cancer ranked as the eighth most common cancer diagnosis and cause of cancer death in women worldwide.⁷⁹ Unlike research on uterine tumors, studies on ovarian tumors using AI frequently incorporate CT and ultrasound (US) in addition to MRI.³ However, this study focuses exclusively on MRI. Ovarian tumors are often large, and their heterogeneity tends to be a problem for quantitative assessment; however, using 3D data to assess the entire tumor may eliminate this problem. To date, various AI models using MRI have been developed for classifying types of ovarian tumors,^80–87 and for predicting the prognosis of high-grade serous carcinoma (HGSC).^88–95

For studies of ML-based models using radiomics data from MRI in multicenter settings for classifying types of ovarian tumors, in 2020, Li et al. developed and validated an ML model using radiomics features from multiparametric MRI to distinguish between epithelial borderline and malignant tumors and compared its efficacy to radiologists’ interpretations. The study included 250 patients from multiple centers for training. The model focusing on solid tumor components outperformed the whole tumor model in the external validation (AUC: 0.90 vs. 0.77), demonstrating its effectiveness in differentiating borderline from malignant tumors with AUCs of 0.91 and 0.92, respectively. The performance of radiologists was significantly lower than that of the model in the internal validation.⁸³ In 2022, Wei et al. conducted a study to evaluate T2WI-based radiomics for preoperatively distinguishing between benign and borderline epithelial ovarian tumor. The study included 309 patients for training. In the external validation, the combined model (radiomics and clinical-radiological models) performed significantly better than the clinical-radiological model (AUCs of 0.86 vs. 0.63, P = 0.021).⁸⁴

In studies of ML-based models using radiomics data from multiparametric MRI in multicenter settings for predicting the prognosis of HGSC, in 2023, Li et al. explored the value of a radiomics nomogram to identify platinum resistance and predict the progression-free survival of patients with advanced HGSC, 301 patients with advanced HGSC underwent radiomics features extraction from the whole primary tumor. The radiomics nomogram, combining radiomics signatures with three clinical characteristics (FIGO stage, CA-125, and residual tumor), had a higher AUC compared with the clinical model alone (0.799 vs 0.747).⁸⁸

To evaluate image classification by DL using the images themselves, in 2021, Wang et al. developed a DL algorithm that leverages a CNN to differentiate between benign and malignant lesions. This was achieved by applying a CNN to T2WI and CE-T1WI images. A dataset comprising 384 lesions was used for training. The study compared the performance of DL (algorithm: EfficientNet) against radiomics models and evaluations made by radiologists. The results showed that, on average, the ensemble DL model had a comparable test accuracy to the radiomics models and outperformed senior radiologists in terms of test accuracy (0.87 vs. 0.74, P = 0.033) and specificity (0.92 vs. 0.70, P < 0.001) while maintaining comparable sensitivity (0.75 vs. 0.83, P = 0.557).⁸⁵ In 2022, Saida et al. compared the effectiveness of a DL approach with that of radiologists for diagnosing epithelial ovarian cancer. This comparison was made using multiparametric MRI in JPEG format. This study included data from 219 patients with malignant tumors in the training set. The CNN (algorithm: Xception) used in this study achieved an AUC ranging from 0.83 to 0.89, demonstrating diagnostic performance on par with that of experienced radiologists.⁸⁶ In 2024, Wei et al. used radiomic methods to assess peritoneal metastasis in patients with epithelial ovarian cancer, utilizing fat-saturated (FS)-T2WI from 297 patients across three centers for model development. This ensemble model, which integrated DL (algorithm: ResNet-50) using images, radiomics features, and clinical data, demonstrated superior diagnostic performance, notably enhancing the AUC and specificity. This model was particularly beneficial for all readers, especially those with less experience. In the external validation sets, the ensemble model achieved the highest AUCs, surpassing both the DL model (0.86 vs. 0.74) and clinical model (0.87 vs. 0.73).⁸⁷

Placenta accreta

The MRI diagnosis of placenta accreta spectrum (PAS) is pivotal in shaping treatment plans, leading to numerous studies in 2024 on the application of AI to diagnose this condition.

For studies of DL model utilizing radiomics features, Peng et al. introduced a T2WI-based DL radiomics model for identifying PAS. The study included 170 pregnant women suspected of having PAS at a single center for model training. This DL radiomics approach outshone traditional clinical and MRI morphologic models, showcasing superior diagnostic capabilities with higher AUCs in 2 external validation sets (0.86 vs. 0.77 and 0.85 vs. 0.68 for clinical models; 0.88 vs. 0.76 and 0.86 vs. 0.78 for MRI morphologic models, respectively).⁹⁶

For image classification by DL using the images themselves, Zong et al. developed a DL model that utilized steady-state acquisition and single-shot fast spin-echo sequences. The model (algorithm: 3D ResNets), trained using data from 227 patients with suspected PAS, employs automatic 3D segmentation to evaluate the risk of adverse pregnancy outcomes associated with placenta accreta. Although the model assessed adverse outcomes with a notable AUC of 0.88, the specific type of imaging used was not disclosed.⁹⁷ Wang et al. created a DL model for the prenatal diagnosis of PAS using T2WI coupled with automatic 3D segmentation. The training dataset comprised 409 pregnant women with suspected PAS from a single center. This fully automatic DL pipeline (algorithm: DenseNet-PP), which focuses on the uteroplacental interface, demonstrated remarkable accuracy, achieving an AUC of 0.90 in external validation and surpassing the diagnostic performance of experienced radiologists; however, the specific type of imaging used was not disclosed.⁹⁸

Segmentation

Segmentation is the task of dividing an image into multiple segments or parts, often to isolate ROIs or distinguish different objects within an image. In medical imaging, segmentation can be used to separate different anatomical structures in a scan, such as organs or tumors, to assist in diagnosis or treatment planning. Though in the field of obstetrics and gynecology, there have been few reports on segmentation using MRI. To assess the performance of the segmentation model, three key metrics were utilized: the dice similarity coefficient (DSC), Hausdorff distance, and average symmetric surface distance. The DSC evaluates the overlap between the predicted segmentation and the ground truth, providing a measure of the accuracy of the model. Hausdorff distance assesses the largest distance between the boundary points of the two segmentations, which helps understand the worst-case error in the spatial extent of the prediction. Finally, average symmetric surface distance calculates the average distance between the surfaces of the predicted and true segmentations, offering insight into the average error across the boundary.⁹⁹ Together, these metrics provide a comprehensive evaluation of the performance of a model in accurately and consistently segmenting images.

In 2021, Kurata et al. developed a CNN to automatically segment EC on MRI. They used 2D sagittal images from T2WI, DWI, and ADC maps with a modified U-net model.¹⁰⁰ In 2021, Hodneland et al. presented a fully automatic approach for segmentation of primary tumor in EC using an axial CE-T1WI oblique 3D volumetric interpolated breath-hold gradient echo sequence with fat saturation and a 3D-CNN (UNet3D). These models can automatically segment EC on MRI and extract radiomics features with high reliability.¹⁰¹ Torrents-Barrena et al. introduced a segmentation approach for intrauterine tissues by employing both radiomics and DL techniques to enhance fetal disorder detection and facilitate surgical planning. The study utilized 60 axial T2WI and 3D US images and applied semiautomatic 3D segmentation to the data collected between 17 and 37 wk of gestation. By integrating selected radiomics features specific to each anatomical structure with data from both MRI and US, the study achieved optimal segmentation performance for fetal and maternal tissues.¹⁰² They also suggested that combining the selected 10 radiomic features per anatomical region along with DeepLabV3+ or BiSeNet algorithms for MRI and PSPNet or Tiramisu for 3D US could lead to the highest fetal/maternal tissue segmentation performance.

In 2023, Mulliez et al. unveiled a DL tool engineered for automated 3D measurements of the uterus using MRI. This tool assessed 900 cases from a single center. The analysis included sagittal and axial T2WI. Impressively, the model (algorithm: VGG-16/VGG-11) demonstrated a key point similarity score of 0.96 within an independent validation composed of earlier cases from the same facility, indicating a high level of precision and a strong alignment with manual measurement techniques.¹⁰³

In 2024, Cui et al. developed a weakly supervised model that used only image-level labels to achieve automatic segmentation of four types of uterine lesions, including EC, uterine leiomyoma, endometrial polyps, and atypical hyperplasia of the endometrium and normal tissues (uterus, endometrium, and cervix) on T2WI. The proposed 2-stage model can be divided into 4 sequential parts: the pixel correlation module, class reactivation map module, inter-pixel relation network module, and DeepLab v3+ module. Images from 196 patients with 4 different types of lesions were used for training, and the proposed model showed good segmentation performance.¹⁰⁴

Image Quality Improvement

Image quality improvement involves techniques and processes aimed at enhancing the visual appearance and clarity of images. This can include increasing resolution, reducing noise, adjusting contrast and brightness, and correcting imperfections such as blurring or distortion. There have been several studies on improving MRI quality in obstetrics and gynecology.

In 2020, Misaka et al. utilized DL (algorithm: U-Net) to enhance the quality of single-shot turbo spin-echo T2WI (SSTSE) of the female pelvis. This study analyzed images from 126 women, employing low-quality images derived from turbo spin-echo T2WI (TSE) as the training input, and TSE were used as ground truth images. During the testing phase, the trained CNN was applied to the SSTSE, resulting in DL-enhanced SSTSE (DL-SSTSE) that exhibited superior quality compared to the original SSTSE. Notably, the DL-SSTSE maintained acceptable quality levels compared with the TSE while preserving the benefits of SSTSE, such as robustness against motion artifacts and efficient acquisition time.¹⁰⁵

In 2022, Tsuboyama et al. assessed the impact of DL reconstruction (algorithm: AIR recon DL, GE Healthcare) and a postprocessing sharpening filter on the image quality of single-shot fast spin-echo T2WI (SSFSE) of the uterus. This study included the parasagittal T2WI of 50 patients and compared images obtained using the periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER), conventional reconstruction (SSFSE-C), DL reconstruction (SSFSE-DL), and DL reconstruction with a sharpening filter (SSFSE-DLF). The findings revealed that SSFSE-DL significantly reduced the contrast between the junctional zone and the myometrium, whereas SSFSE-DLF notably enhanced the image quality of SSFSE of the uterus to a level comparable to that of the PROPELLER sequence.¹⁰⁶

Challenges and Future Prospects

Most AI research delivers superior performance, as evidenced by its higher AUC and accuracy. In particular, radiomics has been shown to match or surpass the diagnostic accuracy of subjective assessments made by radiologists. Radiomics predicts the outcomes of interest better than clinical-radiological models, and in most studies, combined models were the best predictors.

However, several challenges still remain in the field.² Data availability and quality: the success of AI models relies heavily on access to diverse, well-annotated medical datasets, which are often scarce and inadequately prepared, introducing biases and constraints on model efficacy. Lack of standardized evaluation criteria: the absence of uniform evaluation standards for AI research complicates quality assessment and makes it difficult to compare outcomes across studies. Publication bias and reporting: the tendency to publish positive outcomes may distort the perceived effectiveness of AI, creating challenges for objective comparisons with radiologists. Integration into clinical workflows: optimizing AI to work seamlessly within existing clinical practices is essential, demanding further exploration of how AI can effectively complement the roles of radiologists. Black box nature of AI: the opaque nature of many AI algorithms poses significant challenges in comprehending their decision-making processes, which could hinder their acceptance in clinical practice. Consequently, the need to demystify the complexity of DL models has increased, giving rise to a specialized field of AI research known as explainable AI.¹⁰⁷ Limited external validation: the prevalence of single center studies with scarce external validation raises concerns about the overestimation of AI capabilities and the generalizability of the findings. Need for prospective studies and multicenter trials: to ascertain the robustness of AI models, additional prospective studies in real-world clinical settings and multicenter trials are required to evaluate their applicability across diverse populations and environments.

In the field of obstetrics and gynecology, radiomics is at the forefront of AI research, advancing towards studies that utilize 3D data from multiparametric MRI. The integration of DL with clinical data has led to the development of more complex models, which in turn contribute to improvements in performance. Feature extraction is a key process applied after image segmentation and processing. Feature extraction measures the characteristics of the gray levels in segmented tumors.¹⁰⁸ Various methods exist, but the Image Biomarker Standardization Initiative (IBSI) guidelines offer a standardized approach for calculating features with established reference values for 107–169 common features.¹⁰⁹ This helps create a consistent process for radiomics studies. In addition to the IBSI, other guidelines, such as the radiomics quality score and Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis are being developed to ensure the quality and transparency of radiomics research.¹¹⁰ Future studies are expected to adhere to these standards to bridge the gaps in radiomic translation.

Currently, many studies are retrospective and conducted under specific conditions, necessitating further development for clinical application. However, integrating AI into medical imaging enhances diagnostic accuracy and efficiency. By utilizing large datasets from multiple centers and automating image processing, AI’s clinical applicability is significantly increased. AI has the potential to revolutionize how radiologists diagnose diseases, enabling earlier and more precise diagnoses and paving the way for personalized treatment strategies in the future.