P-348 Artificial intelligence-based machine learning to diagnose and classify adenomyosis from ultrasound scans: a model development study

Abstract Study question Can an image recognition model, utilising automated deep learning, be trained to diagnose and classify the severity of adenomyosis on ultrasound? Summary answer The automated deep learning model developed in this study can accurately diagnose and classify the severity of adenomyosis from ultrasound images. What is known already Adenomyosis is prevalent across all stages of life, starting from adolescence to perimenopause, however it remains underdiagnosed. The diagnosis and classification of adenomyosis are essential for correlating with symptom severity and counselling about disease prognosis, management options and long-term impact. Diagnosing and classifying the severity is a time-consuming and complex process that depends on the operator’s experience. Machine learning (ML) has an enormous potential to assist healthcare professionals with image classification tasks. It can directly process and automatically learn mid to high-level abstract features acquired from images using a deep architecture model without requiring the manual definition of features. Study design, size, duration This observational cohort study utilised ultrasound images from four assisted conception units in the United Kingdom from February 2023 to February 2024. Labelled images were used as input for training and testing of the model. The complete dataset was manually split into 9:1 ratio; 90% of all images were used as training dataset (80% for training and 10% for validation) and 10% as test dataset. Ethics approval was obtained from the Institutional Review Board (IRB). Participants/materials, setting, methods An automated ML tool (Vertex AI Vision) was trained and tested to diagnose and classify adenomyosis severity. Two investigators, experienced in gynaecological ultrasound scanning, labelled two and three-dimensional images into four categories: normal uterus, mild, moderate, and severe adenomyotic uterus, as per standardised ultrasound criteria. The labelled images were formatted and uploaded on Google Cloud to train the neural network based on pattern recognition. The model’s performance was determined for an internal test dataset. Main results and the role of chance A total of 3,500 ultrasound images were obtained from 3,194 patients, and the internal test dataset consisted of 350 ultrasound images from 292 patients. An overall and per-category score threshold of 0.5 was adopted to provide a balance between reasonable precision and recall during the evaluation. The investigator’s categorization was used as the reference standard to evaluate the performance of the automated machine-learning model. The model’s overall accuracy for diagnosing and classifying adenomyosis, as determined by the area under the precision-recall curve, was 80.24% across all score thresholds. The overall precision was 83.00%, and recall was 65.75% at a confidence threshold of 0.5. The precision and recall were 91% and 78% for normal uterus, 67% and 53% for mild adenomyosis, 86% and 68% for moderate adenomyosis, and 88% and 64% for severe adenomyosis. The algorithm achieved a high accuracy for the internal dataset. The time taken by a trained healthcare professional to label and process 3,500 images into 4 categories was 70 hours. The time taken by an untrained automated ML tool to get trained, process, and label the images into 4 categories was 4 hours, saving 66 hours per healthcare professional. Limitations, reasons for caution Due to the inherent nature of non-invasive investigations for the population attending the assisted conception unit, histological confirmation of adenomyosis was not possible. In addition, the model’s performance has been tested using an internal test dataset and needs further validation using an external test dataset from a different population. Wider implications of the findings This novel study leverages the power of artificial intelligence to develop an automated machine learning model for diagnosing and classifying adenomyosis from ultrasound scans. This tool can potentially enhance the diagnostic capacity and efficiency of healthcare professionals and empower women by providing better care through early and uniform diagnosis. Trial registration number No