TL;DR
This paper introduces a fully automatic deep learning method for segmenting the pelvic floor levator hiatus in ultrasound images. By employing a novel self-normalizing neural network (SNN) with SELU activation, the approach achieves high accuracy comparable to manual segmentation and outperforms existing U-Net architectures, while also reducing computational demands.
What Problem Does This Paper Address?
Pelvic Organ Prolapse (POP) is a common condition affecting a significant portion of women, characterized by the descent of pelvic organs. Assessing the anatomical integrity and function of the pelvic floor is crucial for diagnosing and managing POP. The levator hiatus, the opening through which POP occurs, is a key anatomical feature. Current methods for measuring the levator hiatus from ultrasound images, while valuable, often require manual segmentation, which is time-consuming and prone to inter-operator variability. Previous semi-automatic methods still necessitate manual landmark identification. This paper aims to address the need for a fully automatic, accurate, and efficient method for levator hiatus segmentation to improve the assessment of pelvic floor disorders.
Methods
The core of this work is a novel fully automatic segmentation method based on a convolutional neural network (CNN). The key innovation lies in the use of a Self-Normalizing Neural Network (SNN), specifically employing the Scaled Exponential Linear Unit (SELU) activation function. SELU is a recently developed activation function that has the unique property of being "self-normalizing," meaning it can maintain a stable mean and variance of activations throughout the network without the need for explicit batch normalization. This offers several advantages:
- Parameter-free: SELU does not require learnable parameters for normalization.
- Mini-batch independent: Unlike batch normalization, SELU's performance is not dependent on the mini-batch size, which can be a significant advantage in medical imaging where datasets can be small, leading to memory constraints.
- Regularization: SELU acts as a form of regularization, potentially leading to more stable training and better generalization.
The authors adapted a U-Net architecture, replacing its standard batch normalization and ReLU layers with SELU activations, creating what they call the SU-Net. They also explored a version with SELU-dropout. The network takes 2D ultrasound images (extracted from 3D volumes) as input and outputs a segmentation mask for the levator hiatus.
Datasets
The study utilized a dataset of 91 ultrasound images extracted from the oblique axial plane (C-plane) of the levator hiatus. These images were acquired from 35 patients and captured at three different stages: Valsalva maneuver, muscle contraction, and rest. Each of the 91 images was manually segmented by three independent operators with at least six months of experience. This dataset was used for training and evaluation using a leave-one-patient-out cross-validation strategy.
Key Insights and Ideas
- SELU as a Game-Changer for Medical Imaging CNNs: The application of SELU in medical imaging CNNs for segmentation is novel. Its self-normalizing properties offer a compelling alternative to batch normalization, especially in scenarios with limited data and memory constraints, which are common in medical imaging research.
- Eliminating Batch Normalization: The success of SU-Net without batch normalization highlights that this standard component might not always be necessary and can even be detrimental in certain contexts, particularly when dealing with small batch sizes.
- Fully Automatic Segmentation is Achievable and Competitive: The paper demonstrates that a fully automatic segmentation method can achieve accuracy comparable to, and in some metrics, superior to manual segmentation, while significantly reducing the need for human intervention.
- Clinical Relevance of Levator Hiatus Segmentation: The study reinforces the importance of accurate levator hiatus measurements for assessing pelvic floor disorders and predicting outcomes like prolapse recurrence.
Key Results
The proposed SU-Net achieved a median Dice similarity coefficient of 0.90 with an interquartile range of 0.08. This performance was found to be equivalent to the manual segmentation performed by the three operators, as indicated by a Williams' index of 1.03.
- Comparison to U-Net: The SU-Net outperformed a standard U-Net architecture (without batch normalization) in terms of Dice, Jaccard, Hausdorff, SMAD, and FPD metrics, with statistically significant differences (p-values < 0.031).
- SELU-dropout: The performance of SU-Net with and without SELU-dropout showed no statistically significant difference.
- Comparison to Other Architectures: SU-Net also showed no statistically significant difference when compared to U-Net+DC and ResNet architectures.
- Inter-operator Variability: The study also quantified inter-operator differences, showing that the automatic method's performance was comparable to the variability observed between human annotators.
Limitations and Future Work
- C-plane Identification: A significant limitation acknowledged by the authors is that the current method requires the manual identification of the C-plane (the specific 2D slice from the 3D ultrasound volume). While the focus was on automating the segmentation within this plane, future work aims to automate the C-plane identification as well.
- Dataset Size: The authors suggest that increasing the number of training images could further improve results, especially for extreme cases and high variability, as the current dataset size is limited.
- End-to-End Solution: The current work is a step towards a minimally interactive workflow. Future plans include developing a more complete analysis pipeline, potentially including automated C-plane identification, optional manual refinement of segmentations, and automated prediction of clinically relevant measurements.
- Clinical Relevance of SELU: While SELU showed promising results and potentially faster convergence, the authors note that drawing quantitative conclusions on the efficacy of SELU units themselves is difficult and warrants further investigation.
Summary updated