Identificazione e classificazione delle lesioni isteroscopiche tramite deep learning: sviluppo dell’algoritmo di rete neurale “HYSTERONET”

Introduction: In clinical practice, hysteroscopy with endometrial biopsy is used for diagnosing abnormal uterine bleeding and endocavitary pathologies. However, the accuracy of the procedure is heavily dependent on the operator's experience. Visual diagnosis during hysteroscopy often has a high rate of false negatives for oncological pathology, especially in postmenopausal women. Objective: This study aims to develop a deep learning (DL) model, named HYSTERONET, for the automated identification and classification of intrauterine lesions from hysteroscopic images. Materials and Methods: A retrospective cohort study was conducted using hysteroscopic images from patients with pathologically confirmed intrauterine lesions. The images were used to build a DL model for lesion identification and classification, with and without the inclusion of clinical factors. Results: A total of 1500 images from 266 patients were analyzed. The best performance for both classification and identification tasks was achieved with the inclusion of clinical factors, resulting in an accuracy of 80.11%, recall of 80.11%, specificity of 90.06%, and F1 score of 80.11% for classification. The identification task achieved an overall detection rate of 85.82%, with a precision of 93.12%, recall of 91.63%, and F1 score of 92.37%. Conclusion: The HYSTERONET model showed moderate performance in the classification and identification of intrauterine lesions. While the inclusion of clinical factors improved the model’s performance, the overall improvement was modest.