[Paper Review] Spontaneous preterm birth prediction using convolutional neural networks

Tomasz Włodarczyk, Szymon Płotka+5 more|arXiv (Cornell University)|Aug 16, 2020

Preterm Birth and ChorioamnionitisMedicine1 citations

TL;DR

This study proposes a one-stage convolutional neural network (CNN) that jointly segments the cervix and predicts spontaneous preterm birth (sPTB) from 2D transvaginal ultrasound images. By extending U-Net with a parallel classification branch, the model achieves a mean Jaccard index of 0.923 ± 0.081 for segmentation and a sensitivity of 0.677 ± 0.042 with a 3.49% false positive rate, outperforming state-of-the-art methods in sPTB prediction.

ABSTRACT

An estimated 15 million babies are born too early every year. Approximately 1 million children die each year due to complications of preterm birth (PTB). Many survivors face a lifetime of disability, including learning disabilities and visual and hearing problems. Although manual analysis of ultrasound images (US) is still prevalent, it is prone to errors due to its subjective component and complex variations in the shape and position of organs across patients. In this work, we introduce a conceptually simple convolutional neural network (CNN) trained for segmenting prenatal ultrasound images and classifying task for the purpose of preterm birth detection. Our method efficiently segments different types of cervixes in transvaginal ultrasound images while simultaneously predicting a preterm birth based on extracted image features without human oversight. We employed three popular network models: U-Net, Fully Convolutional Network, and Deeplabv3 for the cervix segmentation task. Based on the conducted results and model efficiency, we decided to extend U-Net by adding a parallel branch for classification task. The proposed model is trained and evaluated on a dataset consisting of 354 2D transvaginal ultrasound images and achieved a segmentation accuracy with a mean Jaccard coefficient index of 0.923 $\pm$ 0.081 and a classification sensitivity of 0.677 $\pm$ 0.042 with a 3.49\% false positive rate. Our method obtained better results in the prediction of preterm birth based on transvaginal ultrasound images compared to state-of-the-art methods.

Motivation & Objective

• To develop an automated, deep learning-based method for predicting spontaneous preterm birth (sPTB) using transvaginal ultrasound images.
• To overcome the subjectivity and inconsistency of manual ultrasound analysis by gynecologists.
• To integrate cervix segmentation and sPTB classification into a single, end-to-end trainable model.
• To improve prediction accuracy over existing methods relying on handcrafted features like cervical length.
• To provide an objective, quantitative alternative to subjective clinical evaluation of ultrasound images.

Proposed method

• The method employs a modified U-Net architecture with a parallel classification branch for joint segmentation and classification.
• The model is trained on 354 2D transvaginal ultrasound images, including 319 controls and 35 preterm cases.
• A multi-task loss function combines binary cross-entropy for segmentation and classification, optimizing both tasks simultaneously.
• Image preprocessing includes inpainting to remove manual annotations (e.g., yellow/green crosses) using Telea’s algorithm.
• Grad-CAM is used to visualize attention regions, revealing that the network focuses on tissue heterogeneity near the ectocervix for classification.
• Three backbone networks (U-Net, FCN, Deeplabv3) were evaluated, with U-Net selected due to superior segmentation performance.

Experimental results

Research questions

• RQ1 Can a single deep learning model effectively perform both cervix segmentation and spontaneous preterm birth classification on transvaginal ultrasound images?
• RQ2 Does the feature representation learned in the U-Net encoder path contain discriminative information for sPTB prediction?
• RQ3 How does the proposed joint segmentation-classification model compare to state-of-the-art methods in terms of segmentation accuracy and sPTB prediction sensitivity?
• RQ4 Where in the ultrasound image does the model focus its attention during sPTB classification, and is this consistent with clinical knowledge?
• RQ5 Can end-to-end deep learning reduce false positive rates and improve prediction reliability compared to manual or feature-based approaches?

Key findings

• The proposed model achieved a mean Jaccard index of 0.923 ± 0.081 for cervix segmentation, indicating high segmentation accuracy.
• The model demonstrated a sensitivity of 0.677 ± 0.042 for preterm birth classification, with a low false positive rate of 3.49%.
• The model outperformed the state-of-the-art method of [12] in both segmentation (IoU: 0.923 vs. 0.91) and classification (recall: 0.677 vs. 0.596).
• Grad-CAM visualization revealed that the network focuses on the lower segment of the cervix and surrounding tissue heterogeneity, aligning with clinical observations of anterior cervical stroma involvement.
• The model achieved a specificity of 0.951, indicating strong performance in correctly identifying control cases.
• The results suggest that deep features from the U-Net encoder path are highly informative for sPTB prediction, supporting the feasibility of end-to-end learning.