Wednesday, December 4 | 1:40 p.m.-1:50 p.m. | W6-SSOB03-2 | Room S402
This scientific study is said to be the first deep-learning (DL) model to use multisequence MRI on a large cohort for identifying endometriosis.
Out of the Mayo Clinic Artificial Intelligence Lab in Rochester, MN, the research aims to reduce the challenges of diagnosing endometriosis through imaging.
A team of radiologists, radiology residents, and a surgeon evaluated the efficacy of DL tools in enhancing the accuracy of multisequence MRI-based detection of endometriosis. They gathered a patient cohort from their institutional database that was composed of patients with surgically and pathologically confirmed endometriosis between 2015 and 2024.
Steps included collecting gynecologic MRIs three months prior to the diagnostic surgery, creating an age-matched control group that underwent a similar MR protocol but without a diagnosis of endometriosis, and, importantly, analyzing MR images using various convolutional neural network (CNN) architectures. A total of 594 patients were included in the case and control group.
For the analysis, researchers used sagittal T1-weighted (T1) pre- and postcontrast, as well as T2-weighted (T2) MRIs. They allocated 12.5% of the dataset for testing, conducting sevenfold cross-validation on the remainder, according to presenter Mana Moassefi, MD, and co-authors. They noted that two abdominal radiologists and one fellow reviewed a random selection of images and documented their endometriosis detection.
"Our findings indicated the most accurate predictions were obtained from T1 pre- and postcontrast images in our validation sets, employing sevenfold cross-validation across 520 cases," Moassefi and colleagues wrote for the session preview.
The researchers said the DL model achieved average values on seven folds as follows: F1 score of 0.792, area under the receiver operating characteristic curve (AUROCC) of 0.898, sensitivity of 0.854, and specificity of 0.861. In addition, occlusion map visualizations provided insights into the model's focus areas, further confirming the model’s diagnostic accuracy in identifying endometriosis lesions, they said.
With many other data points to be shared during the session, the Mayo Clinic team found that the model showed results nearly equivalent to human detection in identifying endometriosis. The radiologist readers involved achieved an 86.66% accuracy in disease detection, they noted. Attend to learn more and ask questions.
