TLDR: A new study introduces an automated deep learning method for segmenting uterine myomas and surrounding structures from MRI scans. Leveraging the publicly available Uterine Myoma MRI Dataset (UMD) and the nnU-Netv2 framework, the method achieved promising Dice scores across different anatomical structures, including myomas, uterine wall, uterine cavity, and nabothian cysts. This research establishes a robust baseline for future studies, aiming to reduce the manual effort in clinical diagnosis and standardize the assessment of uterine fibroids.
Uterine fibroids, also known as myomas, are very common non-cancerous growths in the uterus, affecting a large percentage of women, especially those of childbearing age. These fibroids can lead to various symptoms like abnormal bleeding, infertility, and pelvic pain. Deciding on the right treatment often depends on the size, number, and exact location of these fibroids.
Magnetic Resonance Imaging (MRI) is a key tool for diagnosing uterine fibroids. However, accurately outlining (segmenting) the uterus and fibroids on MRI scans to measure their volume, shape, and location is a manual process. This manual segmentation is time-consuming, labor-intensive, and can vary significantly between different experts, both before and after treatment. This highlights a clear need for an accurate and automated way to segment uterine fibroids.
Recently, deep learning algorithms have shown remarkable progress in medical image segmentation, often outperforming older methods. These advanced algorithms offer the potential for fully automated segmentation. While several studies have explored using deep learning for automated uterine fibroid segmentation, most have relied on private datasets, making it difficult to validate and compare results across different studies.
A new study addresses this challenge by utilizing the publicly available Uterine Myoma MRI Dataset (UMD). This dataset is a valuable resource, comprising sagittal T2-weighted pelvic MRI scans from 300 patients diagnosed with uterine myoma. It includes detailed, pixel-level annotations for four structures: the uterine wall, uterine cavity, myoma, and nabothian cyst. Collected between 2015 and 2023, the UMD dataset is a high-quality, expert-validated resource for developing and evaluating medical image segmentation algorithms.
A Novel Automated Segmentation Approach
The researchers propose a method based on the nnU-Netv2 framework for fully automated segmentation of the uterine wall, uterine cavity, uterine myoma, and nabothian cyst using the UMD dataset. For their study, MRI scans from 246 patients were used for training the model, with the remaining 54 scans reserved for testing its performance. The method employs a U-Net-based architecture with five resolution levels, designed to automatically configure itself based on the dataset’s characteristics, eliminating the need for manual fine-tuning of many parameters.
The model was trained using a combination of Dice loss and cross-entropy loss, which helps optimize both the overlap between predicted and actual segmentations and the accuracy of classifying each tiny part (voxel) of the image. The performance was primarily evaluated using the Dice similarity coefficient (DSC), which measures how well the predicted segmentation matches the true segmentation.
Also Read:
- RETFound Foundation Model Adapted for Precise Optic Disc Segmentation
- AI Breakthrough in Proton Therapy Planning for Head & Neck Cancers
Promising Results and Future Implications
Despite the inherent complexities of myomas, such as their varied shapes, sizes, and appearances, the proposed method achieved a mean Dice score of 0.70 for myoma segmentation. This indicates its robustness in identifying and outlining myomas across diverse cases. While there was some variability in scores, suggesting challenges with atypical or low-contrast cases, the overall performance is a significant step forward.
Beyond myomas, the method also performed impressively on other structures, achieving Dice scores of 0.86 for the uterine wall, 0.79 for the uterine cavity, and 0.68 for the nabothian cyst. The high consistency in uterine wall segmentation was particularly notable. Overall, the method achieved a mean Dice score of 0.76 across all classes, validating its effectiveness as a powerful framework for uterine MRI segmentation.
This study demonstrates that the proposed method can reliably segment uterine myomas along with surrounding structures using a publicly available MRI dataset. This approach offers a robust baseline for future research and holds significant promise for reducing the clinical burden associated with manual segmentation, leading to more standardized assessments of myoma characteristics. For more details, you can refer to the full research paper.
