EP01.02: Automatic diagnosis of Caesarean scar pregnancy based on ultrasound images by the deep learning model
Year: 2024
Type: conference-abstract
Abstract: To evaluate the performance of four convolutional neural network (CNN) models on early detection of CSP at the early pregnancy stage based on transvaginal ultrasound (TVS) images. Retrospective study of 641 women with gestational age between 4 and 9 weeks who underwent evaluation by TVS between 6/2013-6/2022 at Zhujiang Hospital. This dataset consists of 2095 images, including 956 images from 168 CSP patients, 1139 images from 322 cases without previous Caesarean delivery and 509 images from 151 cases with Caesarean delivery history. The images were included in the training of four CNN models (ResNet50, Swin-Transformer, VGG and MobilenetV2). Additionally, we constructed a fusion model for the independent training test. For validation, 80% of the images from the dataset were utilised for training each model and 20% for testing. Evaluation metrics included F1 score, specificity, accuracy, and AUC. Furthermore, characteristics of CSP images were manually segmented using Matlab software to extract the anatomical boundaries of uterine serosa, gestational sac, and endometrium. From the segmented images, 352 representative images were selected as an independent dataset and incorporated into four deep learning image segmentation models: U-Net, PSP-Net, DeeplabV3 and SegFormer. Each segmentation model used the 80/20 split for training and testing. In the image segmentation models, the outcomes were analysed via the mIoU, mPA, mRecall and mPrecision. In diagnosing CSP, the VGG model achieved the highest AUC value of 0.8618, along with accuracy of 77.3%, sensitivity of 79.8%, specificity of 75.8% and F1 score of 0.803. For the ultrasound image segmentation task in CSP, the U-Net model achieved the mIoU of 0.604, with mPA of 0.729, mRecall of 0.729 and mPrecision of 0.696. The VGG model achieved satisfactory performance in CSP prediction showing its potential to reduce the workload of physicians and improve the accuracy of expert diagnosis of CSP. The segmentation accuracy of U-Net segmentation model need to further improve.
DOI: 10.1002/uog.28017
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4402359810