An Enhanced Deep Learning and Radiomics-based Software Pipeline for Uterine Fibroid Detection in MRI

Abstract

Uterine fibroids, or leiomyomas, are the common benign tumor of the female reproductive system. With uterine fibroid diagnosis, there can be many manifestations may be associated like heavy bleeding, pelvic pain, or pressure, infertility, or even malignant transformation, therefore making it a diagnosis worth reviewing due to its clinical significance. It is important to get an early and correct diagnosis of uterine fibroids, as it will directly affect the management of the condition and potential overall improvement in the patient's outcomes. In this study, here present a reproducible hybrid pipeline that integrates deep learning and radiomics for the automatic detection of uterine fibroids in MRI. The hybrid framework includes a U-Net based segmentation module to locate uterine fibroids, a ResNet-18 based region of interest (ROI) classifier to predict benign vs malignant, and PyRadiomics to extract handcrafted radiomics features. A LightGBM meta-learner was employed to perform feature-level stacking with the deep learning CNN embeddings and radiomics descriptor features. Tested the framework on datasets with uterine myoma data (Uterine Myoma Dataset (UMD)) and TCGA-UCEC. Shown that the hybrid framework achieves better performance than the single-paradigm models with U-Net Dice scores of 0.91 for segmentation and prediction accuracy of 95.1% and AUC of 0.97. In the framework, compared our hybrid setup against CNN-only, radiomic-only, and transformer-based approaches, to show the advantages of stacking. The pipeline was designed using PyTorch, and made the study reproducible by providing open-source scripts available on GitHub, to encourage future research efforts around gynecological AI. The framework has promise in the clinical environment as a supportive decision tool for radiologists and gynecologists that can assist in the early diagnosis of uterine myomas and help in treatment selection.

Ähnliche Arbeiten