AI can be defined as the development of algorithms and systems that enable machines to perform tasks such as learning from data, making autonomous data-driven decisions and adapting to new situations without being explicitly programmed for that.
All AI algorithms require a training phase that can be performed with different methods, such as supervised and unsupervised learning.
In supervised learning, the algorithm is trained with labelled input data, while un-supervised learning algorithm are fed with unlabeled data.
AI encompasses various subfield such as “classical” machine learning and deep learning, which differ in terms of their algorithm structure, data requirement and capabilities. “Classical” machine learning uses linear algorithm to extract features from an image.
Feature extraction is not automatic and classical machine learning requires engineers to design and select relevant features for the model to learn from.
Deep learning (DL) algorithm are designed as multi-layer neural network capable of automatically extract features and learn useful non-linear representations, directly from the raw data.
AI has been employed to improve the accuracy, efficiency, and effectiveness of radiological imaging.
In this setting, AI has shown promising results with various applications in female pelvis MRI, including: - Assisted image acquisition, reconstruction and segmentation AI algorithms, particularly DL techniques, have had substantial success in accelerating MRI acquisition.
Algorithms can reconstruct high-quality images from low-quality acquired MRI data.
Specific deep learning denoising algorithms have been developed to improve signal-to-noise ratio, while other AI techniques have been used to recommend the optimal MRI protocol based on patient history data.
In the image segmentation field, DL algorithms like convolutional neural networks (CNNs) have shown their potential in automating and improving the accuracy of MRI tumor segmentation in cervical, endometrial and ovarian cancers and also in organs at risk (OARs) segmentation for radiotherapy planning of cervical cancer. - AI-powered diagnosis and staging of gynecologic malignancies AI applications in lesion characterization have been focused mainly on differentiating benign from malignant ovarian masses, with promising results.
Regarding endometrial cancer, AI has demonstrated its effectiveness and efficiency in predicting deep myometrial invasion with similar accuracy of the radiologists’.
Moreover, the support of AI has proven its accuracy in predicting lymph-vascular space invasion and lymph node metastases in both endometrial and cervical cancer.
Finally, studies have shown that MRI-based radiomic analysis could help predict endometrial cancer grade and epithelial ovarian cancer subtypes.
All the presented applications represent a substantial improvement from traditional methods and can potentially reduce the time taken for diagnosis and treatment planning. - Treatment response prediction The development of prediction machine learning (ML) algorithms based on radiomics and radiogenomics has shown potential in quantifying treatment response and predicting prognosis.
In cervical cancer, several authors found radiomics to be superior to clinic-pathologic models in predicting response to chemoradiotherapy in locally advanced cervical cancer (LACC).
Currently, the complete response rate is approximately 60-90% for patients with LACC treated with concurrent chemoradiation and intracavitary brachytherapy.
In this context, the early identification of poor responders would permit to adjust therapeutic strategies during treatment. - Prediction of prognosis To date, a large number of studies have analyzed the association between radiomics and prognostic factors, often with good results.
However, the lack of biological validation of these radiomic models prevents them from reaching higher levels of evidence, thereby limiting their application in clinical practice.
AI can facilitate the integration of radiomics, which involves high-dimensional mineable imaging data, with clinical and genetic information, thus creating a more comprehensive patient profile.
This integrated analysis could help validate radiomic models biologically, aligning radiomic features with specific biological characteristics, molecular profiles, or genotypic traits.
This could potentially lead to a deeper understanding of tumor heterogeneity predicting the patterns of recurrence.
In endometrial cancer, radiogenomics can potentially identify unique imaging features and genetic alterations associated with different tumor subtypes, enabling a more precise understanding of individual tumor biology.
The majority of MRI-based models for predicting prognosis were developed for cervical cancer.
The recurrence rate in LACC is 16 - 60% and it is mainly dependent on FIGO stage.
Several researchers have used machine-learning based models to predict overall survival and progression-free survival with good results.
Overall, most studies have concluded that radiomics can predict the aforementioned outcomes better than clinic-radiological-pathological models.
Challenges and Limitations Despite the great potential and success of AI techniques in female pelvis MRI, there are some challenges and limitations to consider.
Deep learning models have demonstrated superior performance over traditional machine learning algorithms.
However, the number of studies employing these models is currently limited.
AI algorithms, especially DL techniques, need an adequately vast and balanced dataset to be trained effectively.
In the medical field, there are not many well-annotated datasets, and the main challenge is to avoid “overfitting” and “biases”.
Overfitting can be defined as the inability of the algorithm to generalize in new unseen data.
Biases generally occur when the data used to train the algorithm are not a representative sample of the general population.
Furthermore, considering that AI algorithms deal with sensitive patient information, ensuring data privacy and security becomes crucial.
Eventually, the lack of transparency and difficult interpretability of AI-based algorithms can discourage radiologists from trusting and validating their results.