A CNN-Transformer Hybrid Network for Recognizing Uterine Fibroids

A severe physical health issue for women is uterine fibroids. Hysteroscopic surgery is an effective way to treat this disease. Considering the strong ability of CNN to gain local features and transformer architecture to obtain global features, this paper proposes a UF classification system (UFCs) using CNN and transformer networks to recognize hysteroscopic images with uterine fibroids for assisting clinicians in the diagnosis of uterine fibroids. The UFCs was trained by a training dataset composed of 9524 images from 240 patients and 199 healthy subjects. Through an evaluation using a test set composed of 2312 images from 33 patients and 36 healthy subjects, the sensitivity, specificity, accuracy, F1-score, precision, and AUC of the proposed method are 94.21%, 83.76%, 88.93%, 89.36%, 84.99% and 0.96, respectively. The proposed method outperformed base models including ConvNeXt, Swin Transformer and Conformer network via testing and comparison, and may be used a computer-aided analysis tool for the recognition of uterine fibroids.