Oral Presentation ESA-SRB-ANZBMS 2024 in conjunction with ENSA

Detecting rectal endometriosis in ultrasounds and magnetic resonance imaging, using artificial intelligence: the IMAGENDO Study (#47)

Jodie C Avery 1 , Yuan Zhang 2 , Steven Knox 3 , Cansu Uzuner 4 , Gustavo Carniero 5 , George Condous 6 , Louise Hull 1
  1. Robinson Research Institute, The University of Adelaide, Adelaide, South Australia, Australia
  2. Australian Insititute of Machine Learning, University of Adelaide, Adelaide, SA, Australia
  3. MRI Department, Benson Radiology, Adelaide, SA, Australia
  4. Acute Gynaecology, Early Pregnancy and Advanced Endosurgery Unit, Nepean Hospital, Kingswood, NSW, Australia
  5. Centre for Vision, Speech and Signal Processing, University of Surrey, Guildford, Surrey GU2, United Kingdom
  6. Ultrasound Department, Omnigynaecare, Sydney, NSW, Australia

In up to 90% of patients with unexplained infertility, endometriosis is found post laparoscopy. In 37% of endometriosis cases laparoscopically diagnosed, bowel endometriosis is found. Currently, there is a 6.4-year delay between first symptoms and surgical diagnosis. Imaging detection of pelvic endometriosis, identifying markers, has a 95% specificity for endometriosis ultrasound (eTVUS) and 72% for magnetic resonance imaging (eMRI). Combining eTVUS and eMRI using Artificial Intelligence (AI), IMAGENDO addresses this delay. Our algorithm development builds on our novel single-modal AI approaches originally for Pouch of Douglas Obliteration Detection, investigating eTVUS and eMRI performance in the automatic detection of rectal endometriosis nodules.

We aimed to develop deep learning (DL) models to automatically classify endometriosis rectal nodules using eTVUS and eMRI. DL models based on a temporal residual network were prospectively trained for each modality, using eTVUS videos and eMRI images. Models were tested on independent test sets and diagnostic accuracies compared to the reference standard sonologist or radiologist classification.

One model is produced for each modality. In a dataset consisting of 519 eTVUS videos, rectal nodules were identified in 35 (6.7%), whereas 484 (93.3%) revealed no nodules. To maintain similar positive/native proportions, enhancing model generalization, we employed stratified 5-fold cross-validation. The model achieved an area under the receiver operating characteristic curve (AUROC) of 85.5% (SD 0.0643). In our eMRI dataset consisting of 127 T2-SPC eMRIs, rectal nodules appeared in 8 (6.4%) images, with 119 (95.6%) images showing no rectal nodules. This model achieved AUROC 69.9% (SD 0.0795).

We have an accurate DL model for eTVUS based rectal nodule classification. However, results using eMRIs alone are imprecise. Future work will combine modalities to improve the eMRI results, allowing extrapolation when either imaging modality is missing. This will enable a faster, more accessible diagnosis for endometriosis, without surgery, provided specialist scanning is available.