JMIR Publications

ABSTRACT

Background:

Transvaginal ultrasound (TVUS) constitutes the first-line medical test for many uterine disorders. However, TVUS requires expertise and experience to diagnose complex conditions like adenomyosis. Whether AI models can help clinicians analyse TVUS videos in real-time remains unanswered.

Objective:

To support gynaecological diagnosis and TVUS analysis, we develop automatic uterine segmentation models performing in real time on TVUS videos.

Methods:

We examine Deep Learning architectures with pre-trained EfficientNet-B7 and InceptionResNetV2 encoder backbones integrated into U-Net and encoder-decoder models, e.g. DeepLabV3 and MAnet. We provide transparency by introducing a two-staged annotation process, an inter-observer variability study, and a reader’s study. By using one multi-centre, public TVUS dataset (Dataset 1) and two single-centre, private TVUS datasets (Dataset 2 and 3), we compare our AI performances to recent studies using this data. We compute three metrics for all segmentation models: Dice score (DSC), intersection-over-union (mIoU), and normalised surface distance (NSD).

Results:

First, our AI models perform similarly or better than recent studies using the same data, e.g. DSC: 0.890±0.021, mIoU: 0.841±0.022, and NSD: 0.502±0.014 for nnUNet v2 on Dataset 2. By exploring inter-observer variability between annotators, we confirm these results with strong spatial concordance and minimal variability in manually segmented volumes (IoU 0.8826 ± 0.0190). Lastly, we show that our AI models can segment uteri in real-time with TVUS videos (Dataset 3), with DeepLabV3 performing best.

Conclusions:

By using three distinct TVUS datasets, we conduct one of the largest studies in automatic uterine segmentation, thus increasing the performance, generalisability and clinical relevance of AI models across TVUS types (2D, 3D, videos), disorders (adenomyosis, niches), and clinical contexts. We prove the decisive role of dedicated Deep Learning architectures, especially DeepLabV3, for the automatic segmentation of various TVUS types. By demonstrating that AI models can segment uteri in real time using unseen TVUS videos, we provide a proof of concept to be tested in real-life TVUS scanning. We conclude that international and interdisciplinary collaborations are needed, and call for more open-source datasets and research funding in this space. Our code is available on [anonymized Github link].