University of Adelaide researchers have identified a new approach for endometriosis classification, believed to be the first of its kind which combines both machine learning models and human knowledge.
The IMAGENDO team from the University's Robinson Research Institute and Australian Institute for Machine Learning (AIML) developed the system -- Human-Artificial Intelligence Collaborative Multi-modal Multi-rater Learning (HAICOMM) -- and published their initial findings in Physics in Medicine and Biology.
HAICOMM was found to eliminate three important challenges in diagnosis by combining AI and human perspectives through several stages.
"First, it uses multi-rater learning to identify a clearer, more reliable label by combining and refining multiple inconsistent or 'noisy' labels for each training sample," says Dr Yuan Zhang, IMAGENDOteam member and grant-funded researcher with the University's Robinson Research Institute.
"Second, it incorporates multi-modal learning, leveraging T1-and T2-weighted MRI images during both training and testing to enhance the system's understanding and accuracy.
"Finally, HAICOMM introduces human-AI collaboration, combining the predictions from clinicians with those of the AI model to achieve more accurate and reliable classifications than either clinicians or AI could achieve alone."
A key diagnostic sign of endometriosis is the obliteration of the Pouch of Douglas (a small space in the female pelvis between the uterus and rectum), which even experienced clinicians can struggle to accurately spot in MRI images.
Research has found that manual classification of Pouch of Douglas obliteration from MRI images has a remarkably high rate of uncertainty, with only 61.4 per cent to 71.9 per cent accuracy.
"This can also complicate the training of reliable AI models," says Dr Zhang.
Endometriosis, where tissue similar to the lining of the uterus grows outside the womb, affects about 14 per cent of individuals assigned female at birth.
It takes on average 6.4 years for patients to receive a formal diagnosis, which generally occurs after the identification of various signs through imaging and/or laparoscopic surgery.
"The long waiting period for a diagnosis lowers the quality of life for those afflicted by the condition and the current reliance on invasive procedures to assist diagnoses escalates healthcare costs, imposing a considerable burden on both healthcare systems and patients," says Dr Zhang.
"These challenges underscore the pressing need for innovative imaging-based diagnostic solutions that can mitigate these issues while enhancing patient care.
"HAICOMM is the first method that explores three important aspects of the challenge in diagnosing endometriosis - multi-rater learning to extract a cleaner label from the multiple 'noisy' labels available per training sample, multi-modal learning to leverage the presence of MRI images for training and testing, and human-AI collaboration to build a system that leverages the predictions from clinicians and the AI model to provide more accurate classification than standalone clinicians and AI models."
The research team will now integrate the technique into the IMAGENDOpatented algorithm, which will leverage both MRI and transvaginal ultrasound images for endometriosis diagnosis.
"By incorporating multi-rater learning, multi-modal MRI data, and human-AI collaboration, we aim to enhance the accuracy and reliability of the IMAGENDO algorithm," says Dr Zhang.
"Additionally, we will expand the system to detect a wider range of endometriosis signs, including bowel nodules, endometriomas, and uterosacral ligament endometriosis, in addition to Pouch of Douglas obliteration.
"The next phase will include evaluating the performance of this enhanced system on diverse datasets and applying it in clinical settings to further assess its practical utility and effectiveness."
