Deep attention networks for the prediction of pouch of Douglas obliteration from transvaginal ultrasound sliding sign videos

Authors Organisations
Type Meeting abstract
Original languageEnglish
Article numberEP34.29
Pages (from-to)448-448
Number of pages1
JournalUltrasound in Obstetrics and Gynecology
Volume54
Issue numberS1
DOI
Publication statusPublished - 30 Sep 2019
Event29th World Congress on Ultrasound in Obstetrics and Gynecology - , Germany
Duration: 12 Oct 201916 Oct 2019
Permanent link
View graph of relations
Citation formats

Abstract

Objectives
Recent studies have demonstrated that the transvaginal sonography sliding sign technique is an accurate method of determining whether the pouch of Douglas (POD) is obliterated in women with suspected endometriosis. In this work we extend earlier work in which state of the art machine learning approaches to image recognition, particularly deep neural networks, were used for automated interpretation of the ‘sliding sign’ videos for automated detection of POD obliteration.

Methods
The data set used for model development consists of 88 ultrasound recordings from different sonographic machines to determine POD obliteration of women presenting with chronic pelvic pain, using the dynamic real‐time ‘sliding sign’ technique. Long term recurrent attention‐based convolutional neural networks were developed for this video classification task. The trained neural networks would take a sequence of frames, and extract spatial‐temporal features hierarchically through multiple layers of information processing, and output scores for the whether the video is interpreted as negative or positive sliding sign.

Results
Our best performing model achieved an accuracy of 80.3%, sensitivity 78.7%, and specificity of 74.7% in hold‐one‐out cross‐validation for predicting POD obliteration.

Conclusions
In spite of the limited dataset, we have demonstrated the potential of using deep attention‐based neural networks for the preoperative prediction of POD obliteration from a number of ‘sliding‐sign’ studies. Further work is needed to improve and evaluate the models using larger datasets.