Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluate Fine-tuning Strategies for Fetal Head Ultrasound Image Segmentation with U-Net

Fangyijie Wang, Guénolé Silvestre, Kathleen M. Curran

TL;DR

The paper addresses the need for accurate fetal head segmentation from ultrasound with limited annotated data by applying transfer learning through fine-tuning a U-Net with a MobileNet V2 encoder. Eight fine-tuning strategies are compared on the HC18 dataset, achieving state-of-the-art-like performance with only 4.4 million trainable parameters, particularly when the decoder is fully unfrozen. Key contributions include a comprehensive evaluation of FT strategies, achieving 97.77% PA, 96.28% Dice, and 92.87% mIoU, and demonstrating that model size can be dramatically reduced without sacrificing accuracy. The work supports deploying compact, TL-based segmentation models in mobile or embedded medical imaging contexts, with code released for replication.

Abstract

Fetal head segmentation is a crucial step in measuring the fetal head circumference (HC) during gestation, an important biometric in obstetrics for monitoring fetal growth. However, manual biometry generation is time-consuming and results in inconsistent accuracy. To address this issue, convolutional neural network (CNN) models have been utilized to improve the efficiency of medical biometry. But training a CNN network from scratch is a challenging task, we proposed a Transfer Learning (TL) method. Our approach involves fine-tuning (FT) a U-Net network with a lightweight MobileNet as the encoder to perform segmentation on a set of fetal head ultrasound (US) images with limited effort. This method addresses the challenges associated with training a CNN network from scratch. It suggests that our proposed FT strategy yields segmentation performance that is comparable when trained with a reduced number of parameters by 85.8%. And our proposed FT strategy outperforms other strategies with smaller trainable parameter sizes below 4.4 million. Thus, we contend that it can serve as a dependable FT approach for reducing the size of models in medical image analysis. Our key findings highlight the importance of the balance between model performance and size in developing Artificial Intelligence (AI) applications by TL methods. Code is available at https://github.com/13204942/FT_Methods_for_Fetal_Head_Segmentation.

Evaluate Fine-tuning Strategies for Fetal Head Ultrasound Image Segmentation with U-Net

TL;DR

The paper addresses the need for accurate fetal head segmentation from ultrasound with limited annotated data by applying transfer learning through fine-tuning a U-Net with a MobileNet V2 encoder. Eight fine-tuning strategies are compared on the HC18 dataset, achieving state-of-the-art-like performance with only 4.4 million trainable parameters, particularly when the decoder is fully unfrozen. Key contributions include a comprehensive evaluation of FT strategies, achieving 97.77% PA, 96.28% Dice, and 92.87% mIoU, and demonstrating that model size can be dramatically reduced without sacrificing accuracy. The work supports deploying compact, TL-based segmentation models in mobile or embedded medical imaging contexts, with code released for replication.

Abstract

Fetal head segmentation is a crucial step in measuring the fetal head circumference (HC) during gestation, an important biometric in obstetrics for monitoring fetal growth. However, manual biometry generation is time-consuming and results in inconsistent accuracy. To address this issue, convolutional neural network (CNN) models have been utilized to improve the efficiency of medical biometry. But training a CNN network from scratch is a challenging task, we proposed a Transfer Learning (TL) method. Our approach involves fine-tuning (FT) a U-Net network with a lightweight MobileNet as the encoder to perform segmentation on a set of fetal head ultrasound (US) images with limited effort. This method addresses the challenges associated with training a CNN network from scratch. It suggests that our proposed FT strategy yields segmentation performance that is comparable when trained with a reduced number of parameters by 85.8%. And our proposed FT strategy outperforms other strategies with smaller trainable parameter sizes below 4.4 million. Thus, we contend that it can serve as a dependable FT approach for reducing the size of models in medical image analysis. Our key findings highlight the importance of the balance between model performance and size in developing Artificial Intelligence (AI) applications by TL methods. Code is available at https://github.com/13204942/FT_Methods_for_Fetal_Head_Segmentation.
Paper Structure (10 sections, 1 figure)

This paper contains 10 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Data Preparation:
  5. Model Design:
  6. Fine-tuning Strategies:
  7. Training and Evaluation:
  8. Experimental Results
  9. Conclusion
  10. Future Work

Figures (1)

  • Figure 1: (a) The first row shows three fine-tuning strategies: U-Net baseline, 0 to 4 layers remain unfrozen within the decoder, and the encoder remains unfrozen. The second row shows three fine-tuning strategies: 0 layer remains unfrozen within the decoder, 0 to 1 layers remain unfrozen within the decoder, and 0 to 2 layers remain unfrozen within the decoder. The last row shows two fine-tuning strategies: 2 to 4 layers remain unfrozen within the decoder, 4 layer remains unfrozen within the decoder. (b) Comparison of our methods with the SOTA methods. (c) Comparison of Pixel Accuracy, Dice Score, and mIoU on Test data set. Mobilenet_v2$\ast$ is the encoder with random weights.