Ansys Papillary Muscle Model

PRECISE FEW-SHOT FAT-FREE THIGH MUSCLE SEGMENTATION IN T1-WEIGHTED

Mri

Sheng Chen⋆, Zihao Tang⋆⋄, Dongnan Liu⋆, Ch´e Fornusek◦, Michael Barnett⋄∨,

Abstract

Precise thigh muscle volumes are crucial to monitor the motor functionality of patients with diseases that may result in vari- ous degrees of thigh muscle loss. T1-weighted MRI is the de- fault surrogate to obtain thigh muscle masks due to its contrast between muscle and fat signals. Deep learning approaches have recently been widely used to obtain these masks through segmentation. However, due to the insufficient amount of pre- cise annotations, thigh muscle masks generated by deep learn- ing approaches tend to misclassify intra-muscular fat (IMF) as muscle impacting the analysis of muscle volumetrics. As IMF is infiltrated inside the muscle, human annotations require ex- pertise and time. Thus, precise muscle masks where IMF is excluded are limited in practice. To alleviate this, we propose a few-shot segmentation framework to generate thigh muscle masks excluding IMF. In our framework, we design a novel pseudo-label correction and evaluation scheme, together with a new noise robust loss for exploiting high certainty areas.

Figure: System Model & Simulation Flow for Ansys Papillary Muscle Model

The proposed framework only takes 1% of the fine-annotated training dataset, and achieves comparable performance with fully supervised methods according to the experimental re- sults.

Index Terms— few-shot, intra-muscular fat, thigh mus-

1. Introduction

Precise thigh muscle volumes are critical for analyzing the motor functionality of patients with diseases that can cause different levels of muscle loss directly or indirectly due to sedentarism, such as amyotrophic lateral sclerosis or multiple sclerosis .

T1-Weighted Mri Is A Commonly

used surrogate for studying muscles as water is their pri- mary mass component and T1-weighted images have good contrast between water and fat . Traditional thigh muscle segmentation approaches usually involve intensity-based al- gorithms and a relatively complex pipeline for pre- and post- processing the images . Recently, with the advancements in deep learning strategies, convolutional neural networks (CNNs) [5, 6, 7] have improved the accuracy of medical seg- mentation tasks and helped reduce human workload signifi- cantly [5, 8]. However, when precise annotations are limited or noisy, the model performance decreases significantly .

In addition, the variability of the acquisition parameters for different scanners and protocols limits the generalization of learning-based methods . Fully-supervised models may overfit a particular dataset and suffer from performance drops on other datasets. Fine-tuning to these other datasets is also constrained by the availability of the scans from these new acquisition setups.

Regarding Thigh Muscle Segmentation,

distinguishing intra-muscular fat (IMF) from thigh muscle remains challenging. IMF is defined as the fat that appears between muscle tissues, and it usually presents a complex structure that is spread into discontinuous branches [11, 12].

Furthermore, muscle regions tend to present intensity inho- mogeneities that disrupt their intensity distribution and affect the segmentation results . Thus, precise human annotation requires expertise and intensive labor, limiting the availability of precise thigh muscle masks where IMF is excluded.

To exploit the limited availability of precise annotations, few-shot segmentation methods commonly use a strategy based on class prototype information . Few-shot methods also address the performance drop problem between different domains by adapting to a new dataset with few scans. Con- sequently, semi-supervised methods have gained increasing attention in medical image segmentation to segment chal- lenging areas and structures for specific clinical tasks .

For our purpose, pseudo-labels derived from precisely la- beled data can enrich the datasets, improving the exclusion of IMF. Several works have shown that combining uncertainty estimation with semi-supervised tasks helps reduce noise in pseudo-labels [15, 16]. There are two popular approaches regarding uncertainty estimation to improve pseudo-labels.

Arxiv:2304.14053V1 [Eess.Iv] 27 Apr 2023

The first approach is to embed dropout layers and use Monte Carlo sampling to generate uncertainty maps to reduce the effect of poorly determined regions. The second approach is to construct multiple decoders with different up-sampling methods [14, 17] and enforce a mutual consistency on the multiple outputs to constrain their uncertainty.

In this paper, we propose a two-stage few-shot segmenta- tion framework specifically designed for precise thigh muscle segmentation excluding IMF. The proposed framework in- cludes a pseudo-label generation (PLG) stage and a pseudo- label denoising (PLD) stage. In the PLG stage, pseudo-labels are generated by MC-Net to enrich the dataset. In the PLD stage, these pseudo-labels are denoised by a novel cor- rection and evaluation strategy. Moreover, a noise-robust loss is proposed to exploit the corrected pseudo-labels. Evaluated by dice for IMF and thigh muscle, our results demonstrate that the proposed method using only 13 (1% of the dataset) precisely labeled images outperforms other methods and can achieve comparable performance to the fully-supervised method.

2. Method

Fig. 1. Overview of the proposed framework. Our framework follows a two-stage process of pseudo-label generation (PLG) and its posterior pseudo-label denoising (PLD).

The overview of our two-stage framework is illustrated in Fig. 1. In the following sections, we focus on the two stages of our approach (PLG and PLD) and our novel loss function, define them formally and present the intuition behind them.

Finally, we introduce our data augmentation scheme to fur- ther extend the training set.

2.1. Pseudo-Label Generation

Considering that thigh muscle annotations usually include the IMF, we chose MC-Net, a semi-supervised approach , as our backbone to combine precise and noisy segmentations to enrich our training set. Specifically, two decoders (θdA and θdB) are designed with different up-sampling strategies based on the features from a single encoder θe. Two probability outputs (PA and PB) are generated by independent decoders using the sigmoid function and then thresholded (> 0.5) into pseudo labels (PLA and PLB). To ensure consistency and reduce uncertainty on the predictions, a cyclical scheme is introduced .

Finally, the loss functions involved in this stage consist of a segmentation loss Lseg and a self-supervised consistency loss Lc. Lseg is only used for training samples with precisely annotated masks ( ˜Y ), where Lc is used for all training data.

(1)

Lc = ||(PA −PLB)||2 + ||(PB −PLA)||2.

(2)

Once the model is trained, pseudo-labels (Ypseudo) for all im- ages (X) in the dataset can be generated by the mean predic- tion between the decoders as follows, where σ is the sigmoid

Function Followed By A Thresholding Operation:

Ypseudo = σ(0.5 × (θdA1(θe1(X)) + θdB1(θe1(X)))). (3)

2.2. Pseudo-Label Correction

Although MC-Net can reduce the uncertainty between the de- coders, the generated pseudo-labels Ypseudo still contain er- rors, especially on challenging samples. To further distin- guish thigh muscle from IMF and reduce the effect of the noisy pseudo-labels, we propose a new denoising strategy.

Inspired by source-free domain adaptation approaches , we argue that the feature within the same category should lie closer to their class prototypes, enforcing a high correlation for samples of the same class. In the case of segmentation, image regions with highly correlated voxels in feature space often have compact intensity ranges .

Thus, Instead

of estimating the distance of a pixel to different class proto- types , we directly exploit the class prototype information hidden in the intensity distribution of a set of given regions.

As the intensity values of the thigh muscle are mostly dis- tributed in the range of (0.2, 0.6) after normalizing the input images (X) to the range (0, 1), we extract coarse masks Ycoarse that exclude most of the IMF tissue. A single coarse

1, Pixel ∈(0.2, 0.6)

0, otherwise.

(4)

Finally, we define the set of corrected pseudo-labels Ycorrected as the intersection between the Ycoarse and Ypseudo masks. Examples are shown in Fig. 2.

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.