Segmenting Fetal Head with Efficient Fine-tuning Strategies in Low-resource Settings: an empirical study with U-Net

Abstract: Accurate measurement of fetal head circumference is crucial for estimating fetal growth during routine prenatal screening. Prior to measurement, it is necessary to accurately identify and segment the region of interest, specifically the fetal head, in ultrasound images. Recent advancements in deep learning techniques have shown significant progress in segmenting the fetal head using encoder-decoder models. Among these models, U-Net has become a standard approach for accurate segmentation. However, training an encoder-decoder model can be a time-consuming process that demands substantial computational resources. Moreover, fine-tuning these models is particularly challenging when there is a limited amount of data available. There are still no "best-practice" guidelines for optimal fine-tuning of U-net for fetal ultrasound image segmentation. This work summarizes existing fine-tuning strategies with various backbone architectures, model components, and fine-tuning strategies across ultrasound data from Netherlands, Spain, Malawi, Egypt and Algeria. Our study shows that (1) fine-tuning U-Net leads to better performance than training from scratch, (2) fine-tuning strategies in decoder are superior to other strategies, (3) network architecture with less number of parameters can achieve similar or better performance. We also demonstrate the effectiveness of fine-tuning strategies in low-resource settings and further expand our experiments into few-shot learning. Lastly, we publicly released our code and specific fine-tuned weights.

• The paper demonstrates that fine-tuning the decoder layers of U-Net enhances segmentation performance with improved Dice and mIoU metrics.
• The study employs a MobileNet v2 encoder to reduce trainable parameters by 85.8%, optimizing the model for resource-constrained environments.
• The approach achieves strong cross-domain generalization, maintaining high accuracy even with limited labeled data from diverse clinical settings.

Efficient Fine-tuning Strategies for Fetal Head Segmentation in Low-resource Settings

The paper presented explores the application of fine-tuning strategies for the segmentation of fetal heads in ultrasound images, specifically utilizing the U-Net architecture in environments characterized by limited data availability. This study addresses a pertinent challenge in ultrasound imaging: the accurate and efficient segmentation of fetal head contours, essential for estimating fetal health parameters like head circumference. The research particularly focuses on optimizing deep learning methodologies to adapt to low-resource settings where labeled data is scarce and computational resources are often restricted.

Methodology Overview

The authors implement a series of fine-tuning strategies leveraging the well-regarded U-Net architecture, known for its proficiency in medical image segmentation. They employ MobileNet v2, a lightweight convolutional neural network, as the encoder due to its reduced computational demand and pre-training on the ImageNet dataset. The study contrasts different fine-tuning strategies, examining both the encoder and decoder components of the U-Net.

A robust experimental framework supports this empirical study, utilizing ultrasound datasets from both high-resource settings (Netherlands and Spain) and low-resource environments (Malawi, Egypt, Algeria). This comprehensive dataset inclusion highlights the method's capability to generalize across diverse clinical environments and imaging devices. The authors evaluate segmentation performance using metrics such as Pixel Accuracy, Dice Similarity Coefficient (DSC), and Mean Intersection over Union (mIoU).

Key Findings

1. Fine-tuning Decoder Layers: The study observes that fine-tuning strategies focused on the decoder layers yield superior performance compared to alternative strategies. This approach results in improved DSC and mIoU, emphasizing the efficacy of adjusting the output layers in improving model accuracy.
2. Parameter Efficiency: The proposed method reduces the number of trainable parameters by 85.8% compared to traditional training approaches without compromising the segmentation performance. This reduction has significant implications in terms of computational cost and resource utilization, making it a viable option for low-resource settings.
3. Data Scarcity Mitigation: By utilizing pre-trained weights and focusing fine-tuning on specific network components, the authors demonstrate that high segmentation accuracy can be achieved even with limited data availability, which is a common constraint in low-resource settings.
4. Cross-domain Transferability: The fine-tuned models exhibit strong generalization capabilities when transferred from high-resource datasets to low-resource datasets. The strategy's success in heterogenous data settings underscores its potential to be deployed widely across different geographical regions with varying levels of healthcare infrastructure.

Implications and Future Work

The findings from this study contribute to the growing body of research focused on enhancing deep learning models' applicability in diverse and resource-constrained environments. By showcasing how effective fine-tuning strategies can enhance segmentation tasks in fetal ultrasound imaging, the research provides actionable insights for both the academic community and medical practitioners aiming to deploy efficient AI solutions in the field.

Future research can build upon these findings by exploring the fine-tuning and deployment of other recent neural network architectures, as well as extending these strategies to other critical medical imaging tasks. Another potential area for investigation is the role of few-shot learning techniques in further reducing data dependency.

In summary, this paper presents a well-evidenced study on utilizing fine-tuning strategies to improve the capacity and efficiency of U-Net for fetal head segmentation, with a particular focus on environments where resources are restricted. The proposed methodologies offer a promising direction for effectively scaling deep learning applications in medical imaging across varied and challenging contexts.