P-289 Optimising the detection of obliterated Pouch of Douglas for endometriosis diagnosis, by combining unpaired endometriosis ultrasounds and magnetic resonance imaging using Artificial Intelligence

Abstract

Abstract Study question Can we improve the diagnostic accuracy of the detection of POD obliteration in endometriosis magnetic resonance imaging, by leveraging results from unpaired eTVUS data sets? Summary answer We illustrate effective multimodal analysis methods improve POD obliteration detection accuracy from eMRI datasets, with an Area Under the Curve (AUC) from 65.0% to 90.6%. What is known already Traditionally, women investigated for pelvic pain and endometriosis, wait 6.4 years for laparoscopic diagnosis. There is a need for a more timely, non-invasive, accessible diagnostic tool. IMAGENDO is designed to combine eTVUS and eMRI using Artificial Intelligence (AI) to address this delay. We have previously demonstrated detection of pelvic endometriosis, including Pouch of Douglas (POD) obliteration, has a 95% specificity from endometriosis ultrasounds (eTVUS) and 72% from endometriosis magnetic resonance imaging (eMRI). This preliminary data has shown our novel multimodal AI approach, using imaging data from eTVUS and eMRIs, can improve diagnostic accuracy when detecting POD obliteration in endometriosis. Study design, size, duration The IMAGENDO study is a program of research designed to create a new diagnostic algorithm for endometriosis. The first part of the study describes the development of our initial algorithm using retrospective cross sectional transvaginal ultrasounds (n = 749), and magnetic resonance images (n = 89 private, n = 8984 public) from 9822 participants overall aged 18 to 45 years, collected between September 2011 and September 2022. Participants/materials, setting, methods Using public MRIs, we pre-trained a machine learning model, and fine-tuned the algorithm using private eMRIs to detect POD obliteration. Then unpaired eTVUSs were introduced to further improve our diagnostic model. We used a machine learning method known as Masked Autoencoder pretraining, which is unsupervised learning reconstructing masked data to generate a larger dataset. Then we embedded the data, compressing a large dataset into a small representation with the most salient features. Main results and the role of chance Scant training samples limited the generalisability of a 3D Vision Transformer to classify POD obliteration from MRI volumes, with an Area Under the Curve (AUC) of 65.0%. However, the masked auto-encoder pre-training partially mitigates this issue, improving the AUC to 87.2%. Adding knowledge distillation, and training a 3D Vision Transformer from scratch on such a small dataset is still challenging, with an AUC of 66.7%. However, adding both together: The knowledge distillation performance of 3D Vision Transformer with masked auto-encoder pre-training reaches an AUC of 77.2%, worse than without knowledge distillation, with an AUC of 87.2%. This could be due to the excessive domain shift between the pre-training dataset and TVUS dataset. However, fine-tuning the model from masked auto-encoder pre-training, the model improves accuracy from AUC=87.2% to AUC=90.6%. With all the steps using unmatched imaging from an alternative modality, this model ultimately demonstrated improvement in the AUC from 65% to 90% on our private MRI dataset. This is the first POD obliteration detection method that distils knowledge from TVUS to MRI using unpaired data, aiming to improve diagnostic accuracy of endometriosis from MRI; and the first machine learning method automatically detecting POD obliteration from MRI data to diagnose endometriosis. Limitations, reasons for caution Our eMRI datasets had some confounding problems, present as a result of artefacts, mislabelling, and misreporting. These were resolved using model checking, student auditing and expert radiology review. We will further test our algorithm with a diagnostic test accuracy study on at least two test cohorts. Wider implications of the findings Pre-training using digital data from different imaging modalities can improve the diagnosis of endometriosis especially when either imaging modality is missing. Provided specialist scanning is available, women with endometriosis will be able to obtain faster diagnosis prior to surgery. Trial registration number ACTRN12623000646640