{"paper_id":"75dd78d9-474b-4a36-a9d0-e7b2b77b85c8","body_text":"Deep Learning-Based Automated Segmentation of\nUterine Myomas\nTausifa Jan Saleem\nDepartment of Computer Vision\nMohamed Bin Zayed University of Artificial Intelligence\nUAE\nMohammad Yaqub\nDepartment of Computer Vision\nMohamed Bin Zayed University of Artificial Intelligence\nUAE\nI. I NTRODUCTION\nUterine fibroids (myomas) are the most common benign\ntumors of the female reproductive system, particularly among\nwomen of childbearing age. With a prevalence exceeding 70%,\nthey pose a significant burden on female reproductive health\n[1]. Clinical symptoms such as abnormal uterine bleeding,\ninfertility, pelvic pain, and pressure-related discomfort play a\ncrucial role in guiding treatment decisions, which are largely\ninfluenced by the size, number, and anatomical location of the\nfibroids [1]. Magnetic Resonance Imaging (MRI) is a non-\ninvasive and highly accurate imaging modality commonly used\nby clinicians for the diagnosis of uterine fibroids. Segmenting\nuterine fibroids requires a precise assessment of both the\nuterus and fibroids on MRI scans, including measurements of\nvolume, shape, and spatial location. However, this process is\nlabor intensive and time consuming and subjected to variability\ndue to intra- and inter-expert differences at both pre- and\npost-treatment stages. As a result, there is a critical need for\nan accurate and automated segmentation method for uterine\nfibroids.\nIn recent years, deep learning algorithms have shown re-\nmarkable improvements in medical image segmentation, out-\nperforming traditional methods. These approaches offer the\npotential for fully automated segmentation. Several studies\nhave explored the use of deep learning models to achieve\nautomated segmentation of uterine fibroids [2]. However,\nmost of the previous work has been conducted using private\ndatasets, which poses challenges for validation and comparison\nbetween studies [2]. In this study, we leverage the publicly\navailable Uterine Myoma MRI Dataset (UMD) [3] to establish\na baseline for automated segmentation of uterine fibroids, en-\nabling standardized evaluation and facilitating future research\nin this domain.\nII. M ATERIALS AND METHODS\nA. Dataset\nUMD [3] is a publicly available dataset comprising sagittal\nT2-weighted pelvic MRI scans from 300 patients diagnosed\nwith uterine myoma. It provides pixel-level annotations of four\nstructures: uterine wall, uterine cavity, myoma, and nabothian\ncyst. Collected between 2015 and 2023, the dataset includes\npatients aged 21 to 86 years (mean±SD: 49.73±12.96), and\nserves as a high-quality, expert-validated resource for devel-\noping and evaluating medical image segmentation algorithms.\nFigure 1 presents a representative example from the UMD\ndataset, showing a sagittal T2-weighted MRI slice and the\nsame slice with overlaid segmentation labels.\nFig. 1. UMD data example (a) Sagittal MRI slice, (b) Sagittal MRI slice with\noverlaid segmentation labels.\nB. Methodology\nWe propose a method based on the nnU-Netv2 framework\n[4] for fully automated segmentation of the uterine wall,\nuterine cavity, uterine myoma, and nabothian cyst using the\nUMD dataset. MRI scans from 246 patients were used for\ntraining, and the remaining 54 scans were reserved for testing.\nThe proposed method uses a U-Net-based encoder-decoder\narchitecture with five resolution levels, where each level\nintegrates convolutional blocks with instance normalization\nand leaky ReLU activation. Downsampling is performed via\nstrided convolutions, while upsampling relies on transposed\nconvolutions to restore spatial resolution. No manual hyper-\nparameter tuning was required, as nnU-Netv2 automatically\nconfigures its architecture and training parameters based on\nthe dataset characteristics.\nWe utilized the 3D full-resolution variant and trained the\nmodel for 400 epochs using a composite loss function com-\nbining Dice loss and cross-entropy loss to jointly optimize\nregion overlap and voxel-wise classification performance. The\ntotal loss Ltotal is defined as:\nLtotal = LDice + LCE. (1)\narXiv:2508.11010v1  [eess.IV]  14 Aug 2025\n\nThe multi-class Dice loss LDice is given by:\nLDice = 1 − 2 PC\nc=1\nPN\ni=1 pi,cgi,c\nPC\nc=1\nPN\ni=1 p2\ni,c + PC\nc=1\nPN\ni=1 g2\ni,c + ϵ\n, (2)\nwhere C is the number of classes, N is the total number of\nvoxels, pi,c is the predicted probability for voxel i and class\nc, gi,c is the ground truth for voxel i and class c, ϵ is a small\nconstant to avoid division by zero. The cross-entropy loss LCE\nis computed as:\nLCE = −\nNX\ni=1\nCX\nc=1\ngi,c log(pi,c). (3)\nThe performance of the model was evaluated using the Dice\nsimilarity coefficient (DSC), also referred to as the Dice score,\nwhich measures the spatial overlap between the predicted and\nground truth segmentations.\nAn example of the segmentation output during inference\non the test set using the proposed method is shown in Fig-\nure 2. The close alignment between the predicted segmentation\nand the ground truth on the test set example highlights\nthe effectiveness of the method in automated uterine MRI\nsegmentation.\nFig. 2. Uterine MRI segmentation during inference on the test data.\nIII. R ESULTS AND DISCUSSIONS\nDespite the inherent challenges posed by myomas, includ-\ning their variable shape, size, and appearance, the proposed\nmethod achieved a mean Dice score of 0.70 for this class\n(Table I), highlighting its robustness in detecting and delineat-\ning myomas across diverse cases. The large standard deviation\nreflects substantial case-to-case variability, as also visualized\nin Figure 3 where some cases exhibit Dice scores below 0.50,\nwhile some others exceed 0.80. This suggests that while the\nmodel can segment myomas accurately in favorable cases,\nchallenges remain for cases with atypical morphology or low\ncontrast. Such variability underscores the importance of further\nrefinements, such as tailored post-processing or additional\ntraining data focused on difficult cases.\nBeyond myoma segmentation, the proposed method also\nperformed impressively across other classes. It achieved Dice\nscores of 0.86 for the uterine wall, 0.79 for the uterine cavity,\nand 0.68 for the nabothian cyst (Table I), demonstrating the\nmethod’s ability to generalize across multiple classes. Notably,\nuterine wall segmentation was highly consistent across cases,\nas reflected by its low standard deviation and compact box\nplot distribution (Table I, Figure 3). In contrast, the lower and\nmore variable scores for the nabothian cyst reflect expected\nchallenges associated with its small size. Overall, the pro-\nposed method achieved a mean Dice score of 0.76 across all\nclasses, validating its effectiveness as a powerful framework\nfor uterine MRI segmentation. Crucially, its performance on\nmyoma segmentation establishes a significant step forward\nin the development of automated tools for managing one of\nthe most prevalent and burdensome conditions in women’s\nreproductive health.\nTABLE I\nDICE SCORES (MEAN ± STANDARD DEVIATION ) ON THE TEST SET .\nLabel Dice Score\nUterine Wall 0.86 ± 0.05\nUterine Cavity 0.79 ± 0.10\nMyoma 0.70 ± 0.27\nNabothian Cyst 0.68 ± 0.38\nFig. 3. Dice Scores for uterine wall, uterine cavity, myoma and nabothian\ncyst on the test set.\nIV. C ONCLUSIONS\nIn this study, we addressed the need for an automated\nand reproducible method for uterine myoma segmentation\nby leveraging a publicly available MRI dataset. The results\ndemonstrate that our proposed method can reliably segment\nuterine myomas along with surrounding structures, offering\na robust baseline for future research. This approach holds\npromise for reducing the clinical burden associated with\nmanual segmentation and enabling more standardizeded as-\nsessment of myoma characteristics.\nREFERENCES\n[1] M. M. McWilliams and V . M. Chennathukuzhi, “Recent advances in\nuterine fibroid etiology,” in Seminars in reproductive medicine , vol. 35,\nno. 02. Thieme Medical Publishers, 2017, pp. 181–189.\n[2] A. Tinelli, A. Morciano, R. Sparic, S. Hatirnaz, L. E. Malgieri, A. Mal-\nvasi, A. D’Amato, G. M. Baldini, and G. Pecorella, “Artificial intelligence\nand uterine fibroids: A useful combination for diagnosis and treatment,”\nJournal of Clinical Medicine , vol. 14, no. 10, p. 3454, 2025.\n[3] H. Pan, M. Chen, W. Bai, B. Li, X. Zhao, M. Zhang, D. Zhang, Y . Li,\nH. Wang, H. Genget al., “Large-scale uterine myoma mri dataset covering\nall figo types with pixel-level annotations,” Scientific Data, vol. 11, no. 1,\np. 410, 2024.\n[4] F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein,\n“nnu-net: a self-configuring method for deep learning-based biomedical\nimage segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021.","source_license":"CC0","license_restricted":false}