Evaluating Deep Learning Architectures for Actin Filament Segmentation under Varying Noise Conditions in Simulated Cryo-ET Tomograms

Abstract

Actin filaments are fundamental components of the cytoskeleton, essential for maintaining cell shape, enabling motility, and facilitating intracellular transport. Cryo-electron tomography (cryo-ET) enables nanometer-resolution visualization of filament networks in situ; however, low signal-to-noise ratios, missing wedge artifacts, and complex 3D architectures present significant challenges for accurate analysis. Manual filament annotation is highly resource-intensive and prone to variability, underscoring the need for automated approaches. In this study, we develop and evaluate deep learning–based semantic segmentation architectures for accurate segmentation of actin filament networks in cryo-ET data. We systematically assess several architectures—3D U-Net, Attention 3D U-Net, TransUNet 3D, and UNETR—on simulated tomograms with known ground truth. Model performance is quantified using Dice scores and Intersection over Union (IoU) to evaluate segmentation performance under challenging imaging conditions. Our results show that no single deep learning architecture consistently outperforms others, highlighting the importance of accounting for filament arrangement and noise characteristics. By providing a comparative evaluation of these architectures, we demonstrate their effectiveness in detecting filamentous structures and offer guidance for future efforts to improve segmentation performance.

I. Introduction

Actin filaments are fundamental components of the cytoskeleton, playing crucial roles in maintaining cell shape, enabling motility, and supporting intracellular transport. Accurate modeling and reconstruction of actin filament networks are essential for understanding their structural organization and functional dynamics across diverse cellular contexts. Disruptions in actin filament organization have been implicated in various disease processes, including neurodegeneration, cancer metastasis, and pathogen invasion. Detailed knowledge of filament geometry, connectivity, and spatial arrangement provides insights into mechanotransduction and intracellular movement.

Cryo-electron tomography (cryo-ET) is a powerful technique for visualizing actin filament networks in their native cellular context at nanometer resolution. However, the tomograms often suffer from low signal-to-noise ratios and missing wedge artifacts. The missing wedge arises because physical constraints prevent tilting the sample through the full 180°, leaving certain viewing angles unrecorded. These limitations obscure filament boundaries and connectivity, making it difficult to accurately reconstruct filament geometry, spatial arrangement, and interactions—information essential for understanding cellular processes.

The manual annotation of actin filaments in cryo-ET data is highly resource-intensive and prone to variability. Accurately tracing individual filaments in three dimensions requires meticulous examination of each tomogram slice by structural biology experts, a process that can take hours to days per dataset depending on its size and complexity. Dense filament packing, inherently low signal-to-noise ratios, and missing wedge artifacts—which arise from incomplete angular sampling during data acquisition—further complicate consistent annotation, resulting in significant inter- and intra-annotator variability. These challenges make manual segmentation laborious, time-consuming, and difficult to standardize across studies. Consequently, there is a need for automated approaches that can reliably segment and reconstruct filament networks, thereby reducing annotation time, improving reproducibility, and enabling high-throughput analysis.

Given the labor-intensive and variable nature of manual annotation, there is growing interest in automated approaches for segmenting actin filaments in a cryo-ET volume. Deep learning-based semantic segmentation has emerged as a transformative method for pixel-wise classification across domains such as medical and biological imaging. However, cryo-ET presents unique challenges: the scarcity of annotated 3D tomograms, combined with imaging limitations such as low signal-to-noise ratios and missing wedge artifacts, often constrains segmentation performance. Consequently, despite advances in deep learning architectures, the evaluation of their comparative effectiveness for segmenting actin filaments in cryo-ET remains largely unexplored.

In this study, we address these research gaps by applying five deep learning–based semantic segmentation architectures—namely 3D U-Net (Çiçek et al., 2016 [1]), Attention 3D U-Net (Nodirov et al., 2018 [2]), TransUNet 3D (Chen et al., 2021 [3]), UNETR (Hatamizadeh et al., 2022 [4]), and SegFormer3D (Perera et al., 2024 [5])—to cryo-ET tomograms of actin filaments. We evaluate the models on two sets of simulated tomograms (Dataset-1 and Dataset-2) with known filament positions and systematically assess each model’s ability to segment filamentous structures under a variety of conditions, considering differences in filament geometry and size.

Among the evaluated architectures, no single model consistently outperforms the others; performance varies based on the dataset and noise level. For instance, 3D U-Net and Attention 3D U-Net achieve the best results in Dataset-1 under different noise conditions. In contrast, in Dataset-2, the transformer-based TransUNet 3D demonstrates superior segmentation performance across most noise levels. Regarding consistency and reliability, Attention 3D U-Net performs most consistently within Dataset-1, whereas TransUNet 3D is the most reliable for Dataset-2. However, this consistency does not hold uniformly across both datasets.

II. Related work

A. Classical Tracing Approaches for Actin Filaments in Cryo-ET

Early efforts to segment and trace actin filaments in cryo-ET maps relied entirely on classical image processing and algorithmic search techniques. The foundational approach, exemplified by Rigort et al. [6], combined cylindrical template matching to accurately determine local filament orientation and similarity with a subsequent path-finding algorithm to connect these segments into continuous centerlines, enabling quantitative analysis of network metrics such as length and branching. Other template-based methods followed, including VolTrac (Rusu et al., 2012 [7]), which employed a stochastic genetic algorithm to search the volume with a filament template.

Loss et al. [8] developed a framework for the automatic segmentation and analysis of filamentous objects in 3D electron tomography that allows the quantification of filamentous networks in terms of their compositional and morphological features. Their proposed approach consists of three steps: (i) local enhancement of filaments by Hessian filtering; (ii) detection and completion (e.g., gap filling) of filamentous structures through tensor voting; and (iii) delineation of the filamentous networks. The authors first validated the approach using a set of specifically designed synthetic data. Afterward, they applied the segmentation framework to tomograms of plant cell walls undergoing different chemical treatments for polysaccharide extraction.

Kovacs et al. [9] developed a template-based deconvolution method to correct the missing-wedge artifacts prior to the tracing; however, this approach required expensive numerical techniques (up to a week of run time) and was likely to interpret the high-density voxels in entire tomogram (including membranes, other biomolecules, and noise) as filaments, leading to false-positive predictions.

BundleTrac (Sazzed et al., 2018 [10]), a semi-automatic method developed to successfully model highly ordered, tightly packed filament structures found in hair cell stereocilia.

These methods established the initial framework for automated cytoskeletal analysis, demonstrating that computational techniques could achieve reproducible tracing that was previously limited to labor-intensive manual efforts. As the field progressed, researchers developed more advanced algorithmic solutions. The Spaghetti Tracer (Sazzed et al., 2022 [11]) framework was designed to efficiently trace semi-regular, loosely aligned bundles by accumulating path densities through dynamic programming from local seed points, assuming that the filaments shared a mean direction. Conversely, Struwwel Tracer (Sazzed et al., 2023 [12]) was developed to address highly irregular, randomly oriented filament networks often found in cellular protrusions. This method is agnostic to any global direction, instead accumulating path densities from multiple seed points in various orientations before iteratively fusing short candidate segments into complete, curved traces based on local collinearity. Despite these advances in robustness, all classical approaches remained fundamentally constrained by their high computational cost—often requiring days to execute—and by the need for manually tuned userdefined parameters.

B. Deep Learning-Based Segmentation Approaches for Actin Filaments in Cryo-ET

With advances in deep neural networks (DNNs), in recent years, a limited number of studies have applied DNN-based models for segmenting actin filaments in cryo-ET data. However, these models are generally designed as generic tools for segmenting a variety of biological macromolecules, rather than being specifically tailored to actin filaments. For example, Chen et al. [13] introduced a convolutional neural network (CNN)-based model for segmentation, which was integrated into the EMAN framework [14]. These models demonstrated the ability of deep neural networks to automatically learn complex spatial and textural representations from volumetric data, improving segmentation accuracy and consistency compared to classical image processing approaches. More recently, the shift toward global feature learning has enabled transformer-based models, such as TARDIS (Transformer-based Rapid Dimensionless Instance Segmentation) [15], which can annotate both membranes and filaments in cryo-ET. TARDIS provides instance-level segmentation and shows superior accuracy over standard U-Net models even with limited data. Nonetheless, it remains a generalized cryo-ET segmentation tool and was not specifically designed for actin filament tracing.

Despite these advancements, a comparative evaluation of deep learning architectures specifically for actin filament segmentation in cryo-ET data has not yet been performed. This study addresses this gap by systematically evaluating the performance of popular deep learning-based semantic segmentation architectures—including 3D U-Net, Attention 3D U-Net, TransUNet 3D, UNETR, and SegFormer3D—on simulated cryo-ET datasets, providing a comprehensive benchmark for future research

III. Simulated cryo-et tomogram

Experimental cryo-ET tomograms of actin filaments with corresponding ground truth annotations are rarely publicly available. The underlying tasks—including image acquisition, 3D tomogram reconstruction—are technically demanding, resource-intensive, and time-consuming. Moreover, manual annotation is prone to variability and bias due to low signal-to-noise ratios, missing wedge artifacts, and the dense packing of filaments. In addition, accurately labeling individual filaments in 3D requires meticulous inspection by structural biology experts. These challenges motivate the use of simulated tomograms to provide reliable references for method evaluation.

We use realistic simulated tomograms generated by TomoSim [16], a cryo-ET simulation tool, based on existing filament core tracings. These simulations reproduce the noise characteristics and missing wedge artifacts commonly observed in experimental cryo-ET maps. The input model for generating the simulated data can be derived through coarse-grained manual annotations or predictions from existing modeling tools on experimental tomograms. Since the model is used only to generate simulated tomograms, it does not need to be exact; as long as it reflects and preserves the native actin filament geometry, it is sufficient.

We generate two sets of realistic simulated tomograms representing diverse filament geometries, as described below.

Semi-regular actin bundle (stereocilia):

This structure corresponds to the tapered regions of stereocilia, where actin filaments are organized into tightly packed, semi-parallel bundles. Filaments exhibit uniform polarity, with barbed ends oriented toward the tip, and the bundle gradually tapers in diameter toward the distal end. Inter-filament spacing is approximately 12–15 nm, reflecting the highly ordered crosslinking observed in vivo.

The ground truth model was generated through manual annotation by a domain expert on a highly noisy experimental tomogram [10], [17]. A set of tomograms with varying noise levels (0.4, 0.6, 0.8, and 1.0) was then generated (for details, see [11]). Each tomogram has a size of 347×639×292 voxels. These tomograms are collectively referred to hereafter as Dataset-1.

Actin network (filopodia):

This structure represents the dense, branched actin filament network characteristic of filopodia. Filaments exhibit mixed polarity and are interconnected via crosslinking proteins, forming a dynamic, mesh-like network. Filament lengths are heterogeneous, and branching occurs at characteristic angles (≈ 70°), as mediated by Arp2/3-like nucleation patterns, resulting in a less regular but mechanically robust architecture.

For this, the ground truth model was generated from the prediction of an automated filament modeling tool [6] applied to a low-quality and highly noisy experimental tomogram. A set of five tomograms with varying noise levels (0.35, 0.50,0.65, 0.80, and 0.95) was then generated (for details, see [12]). Each tomogram has a size of 200×200×71 voxels. These tomograms are collectively referred to hereafter as Dataset-2.

IV. Methodology

A. DNN architecture for segmentation

We use five semantic segmentation architectures: 3D U-Net ([1]) as a classic baseline, and four modern architectures—Attention 3D U-Net [18], TransUNet 3D [3], and UNETR [4] and SegFormer3D [5] — representing convolution-based, hybrid, and transformer-based designs. This selection enables a comprehensive evaluation across different segmentation paradigms.

3D U-Net:

The 3D U-Net architecture [1] follows a symmetric encoder–decoder design with skip connections that concatenate high-resolution feature maps from the contracting path with those from the expansive path. This design enables effective integration of contextual and spatial information, supporting precise localization in segmentation tasks. 3D U-Net has been widely applied in biomedical and natural image segmentation, demonstrating strong performance across various semantic segmentation benchmarks ([19], [20], [21]).

Attention 3D U-Net:

The Attention 3D U-Net [18] extends the original U-Net (Ronneberger et al., 2015 [22]) by incorporating learnable Attention Gates (AGs) at the skip connections. These gates generate soft, region-specific weights that suppress irrelevant background features while enhancing target-relevant activations. This adaptive attention mechanism improves feature selection and focus, which has led to successful applications in biomedical image segmentation tasks, particularly in settings with low contrast or high noise [23], [24].

TransUNet 3D:

TransUNet 3D [3] integrates convolutional and Transformer-based components to overcome the limited receptive field of CNNs. A CNN front-end extracts local spatial features, which are subsequently processed by a Vision Transformer to capture global dependencies through Multi-Head Self-Attention (MSA). This hybrid design combines detailed local information with broader semantic context. TransUNet 3D has been successfully applied to a range of medical and natural image segmentation tasks, achieving strong performance in modeling complex spatial relationships.

UNETR:

The UNETR architecture [4] is designed for 3D volumetric segmentation using a pure Transformer encoder. The input volume is divided into non-overlapping 3D patches, which are linearly projected into a token sequence and processed through the Transformer to learn global volumetric representations. Multi-scale features extracted from the Transformer layers are connected via skip connections to a convolutional decoder, enabling precise spatial reconstruction. UNETR and its variants have shown competitive performance in 3D medical image segmentation tasks, leveraging self-attention to model long-range spatial dependencies [25].

SegFormer3D

SegFormer3D [5] is a lightweight and efficient hierarchical Transformer designed for 3D volumetric medical image segmentation. The architecture incorporates an overlapping patch-merging technique within a multi-stage hierarchical encoder to capture multi-scale volumetric features while maintaining spatial continuity between voxels. It employs an efficient self-attention mechanism that substantially reduces the computational complexity associated with 3D volumes. A key distinction lies in its lightweight all-MLP (Multi-Layer Perceptron) decoder, which aggregates local and global features from the encoder stages to generate accurate segmentation masks efficiently, thereby avoiding the computational overhead of complex convolutional decoders. This design makes SegFormer3D a resource-efficient yet competitive alternative to larger state-of-the-art models in medical imaging.

B. Training, Validation, and Testing Split

Following standard practice in segmentation tasks, each tomogram was divided into training, validation, and testing subsets. The training set (approximately 40%–50% of Z-slices, depending on the tomogram) was used for model optimization, the validation set (20%-30%) for hyperparameter tuning and overfitting control, and the testing set (20%-30%) for final evaluation. All subsets preserved comparable distributions of filament and background regions, as well as consistent noise characteristics.

C. Description of Training Process

The training process was carried out using multiple segmentation architectures, including 3D U-Net, Attention 3D U-Net, TransUNet 3D, UNETR, and SegFormer3D.

Input and Output Specifications

• Patch size: 32 × 32 × 32 voxels for simulated Dataset-1 with stride 32 and 16 × 16 × 16 voxels for simulated Dataset-2 with stride 16.

• Number of channels: 1 (grayscale input).

• Output classes: Two classes, background and filament (foreground).

• Label encoding: Binary encoding was used, where the voxel values of 0 represent the background and 1 represent the filament structures.

Network Architecture:

We followed the default settings provided by the original authors of these semantic segmentation architectures.

1. 3D U-Net & Attention 3D U-Net:

   Encoder: Four stages with Conv3DBlocks (3 × 3 × 3 kernels, padding=1) and MaxPool3D downsampling.

   Channels: Features→[32, 64, 128, 256, 512] across encoder levels.

   Decoder: Symmetric upsampling with ConvTranspose3D.

2. TransUNet 3D:

   Encoder: Three CNN stages with 2th downsampling.

   Channels: Features→[32, 64, 128, 256].

   Transformer: Twelve blocks, embed dim=768, num heads=12.

   Decoder: Three stages with ConvTranspose3D.

3. UNETR:

   ViT backbone: Hidden size=768, twelve transformer blocks, num heads=12.

   Encoder branches: Features→[64³, 32³, 16³, 8³].

   Decoder: Four stages using ConvTranspose3D.

   Feature channels: [16, 32, 64, 128].

4. SegFormer3D:

   Architecture: Four-stage hierarchy with hierarchical patch embeddings:

   • Stage 1: kernel=7, stride=4.
   • Stages 2 to 4: kernel=3, stride=2.

   Channels: embed dims=[32, 64, 160, 256].

   Spatial reduction ratios: [8, 4, 2, 1].

   Decoder: MLP projections, concatenation, 3 × 3 × 3 fusion conv, 4 × 4 trilinear upsampling.

Training Hyperparameters:

The models were trained using the following hyperparameter configurations:

• Loss function: Dice loss.

• Optimizer: Adam optimizer.

• Learning rate: Initialized at 0.001 and adjusted using the ReduceLROnPlateau scheduler in minimization mode, which reduces the rate by a factor of 0.1 with a patience of 25 epochs.

• Batch size: 16.

• Maximum Number of epochs: 250.

• Early stopping: Applied with a patience of 25 epochs based on validation loss.

• Regularization: Dropout with a rate of 0.3.

Hardware and Software:

Experiments were conducted on different GPU configurations according to model requirements, including NVIDIA A100 (40 GB VRAM) and T4 (16 GB VRAM) GPUs on Google Colab, as well as an RTX 5000 GPU (32 GB VRAM) on a local workstation. All models were implemented in PyTorch (version 2.7.1).

V. RESULTS AND DISCUSSION

A. Evaluation Metrics

To evaluate the DL model’s ability to segment filamentous structures under different noise conditions, we computed the Dice score (F1 score) separately for the background and filament classes, and subsequently calculated both micro- and macro-averaged Dice scores. The Dice score is a widely adopted semantic segmentation metric that quantifies the overlap between a prediction and the ground truth [26], [27]. It provides a balanced assessment by combining precision (the proportion of predicted filament voxels that are correct) and recall (the proportion of true filament voxels that are correctly identified). This approach allows us to quantify both the overall segmentation performance and the model’s specific capability to accurately delineate fine filamentous structures in noisy cryo-ET tomograms.

The Dice score is defined as:

$$\text{Dice}=\frac{2\times |P\cap G}{\left|P\right|+\left|G\right|}$$ (1)

where P and G denote the sets of predicted and ground truth voxels, respectively. A higher Dice score indicates stronger agreement between the prediction and the reference annotation.

To further assess segmentation quality, we also used the Intersection over Union (IoU), or Jaccard Index [28], [29], defined as:

$$\text{IoU}=\frac{|P\cap G|}{|P\cup G|}$$ (2)

IoU provides a stricter measure of spatial overlap by evaluating the ratio of intersection to union between the ground truth and predicted segmentation. While the Dice score emphasizes the harmonic mean of precision and recall, IoU offers a complementary perspective that is particularly useful for assessing fine structural segmentation such as filament networks.

B. Quantitative Results

To evaluate the effectiveness of the various deep learning models, we conducted a comprehensive quantitative analysis on the two actin datasets described earlier (Dataset-1: Semi-regular Actin Bundle and Dataset-2: Actin Network). The following sections present detailed comparisons of segmentation performance of various models across multiple noise levels for each dataset.

Performance on Dataset-1:

Table I summarizes the segmentation performance of five deep neural network models—3D U-Net, Attention 3D U-Net, TransUNet 3D, UNETR, and SegFormer3D—across four noise levels (0.40, 0.60, 0.80, and 1.0). At low noise (0.40), Attention 3D U-Net achieved the highest Dice (0.875) and IoU (0.778) scores for the filament class, slightly outperforming 3D U-Net. As noise increased, all models exhibited a decline in performance. At moderate noise (0.60), 3D U-Net performed best with a Dice of 0.770 and IoU of 0.626. For higher noise levels (0.80 and 1.0), 3D U-Net demonstrated superior robustness, maintaining higher segmentation accuracy than the other architectures. Overall, Attention 3D U-Net yielded the best results at lower noise levels, whereas 3D U-Net showed greater stability and resilience under higher noise conditions.

TABLE I:

Comparison of segmentation performance (Dice and IoU) of five deep learning architectures on Dataset-1 (semi-regular actin bundle) under different noise conditions

Noise level	Model	Metric	Background	Filament	Micro (Overall)	Macro (Class Avg.)
0.40	3D U-Net	Dice	0.998	0.875	0.998	0.936
		IoU	0.996	0.777	0.996	0.887
	Attention 3D U-Net	Dice	0.998	0.875	0.998	0.937
		IoU	0.996	0.778	0.996	0.887
	TransUNet 3D	Dice	0.998	0.845	0.998	0.921
		IoU	0.995	0.732	0.995	0.864
	UNETR	Dice	0.997	0.814	0.997	0.906
		IoU	0.995	0.687	0.995	0.841
	SegFormer3D	Dice	0.997	0.740	0.997	0.868
		IoU	0.993	0.587	0.993	0.790
0.60	3D U-Net	Dice	0.997	0.770	0.997	0.883
		IoU	0.993	0.626	0.993	0.810
	Attention 3D U-Net	Dice	0.997	0.762	0.997	0.879
		IoU	0.993	0.615	0.993	0.804
	TransUNet 3D	Dice	0.996	0.729	0.996	0.863
		IoU	0.992	0.574	0.992	0.783
	UNETR	Dice	0.995	0.662	0.995	0.829
		IoU	0.991	0.495	0.991	0.743
	SegFormerSD	Dice	0.995	0.582	0.995	0.788
		IoU	0.990	0.410	0.990	0.700
0.80	3D U-Net	Dice	0.993	0.518	0.993	0.756
		IoU	0.987	0.350	0.987	0.668
	Attention 3D U-Net	Dice	0.994	0.625	0.994	0.810
		IoU	0.988	0.454	0.988	0.721
	TransUNet 3D	Dice	0.987	0.390	0.987	0.689
		IoU	0.974	0.243	0.974	0.608
	UNETR	Dice	0.993	0.502	0.993	0.748
		IoU	0.986	0.335	0.986	0.661
	SegFormer3D	Dice	0.994	0.398	0.994	0.696
		IoU	0.988	0.249	0.988	0.618
1.0	3D U-Net	Dice	0.994	0.504	0.994	0.749
		IoU	0.988	0.337	0.989	0.662
	Attention SD U-Net	Dice	0.993	0.501	0.993	0.747
		IoU	0.987	0.331	0.987	0.660
	TransUNet 3D	Dice	0.993	0.230	0.993	0.611
		IoU	0.986	0.130	0.986	0.558
	UNETR	Dice	0.993	0.287	0.993	0.640
		IoU	0.983	0.168	0.985	0.577
	SegFormer3D	Dice	0.991	0.201	0.991	0.596
		IoU	0.981	0.112	0.981	0.547

Performance on Dataset-2:

Table II presents the segmentation results for Dataset-2 across five noise levels (0.35–0.95). At low to moderate noise level (0.35 and 0.50), TransUNet 3D achieved the best performance, yielding Dice scores of 0.673 and 0.602 for the filament class. At 0.65 noise level, Attention 3D U-Net slightly outperformed the other models with a Dice score of 0.541, while at very high noise levels (0.80 and 0.95), TransUNet 3D and UNETR demonstrated stronger robustness, maintaining filament Dice scores above 0.47. Overall, TransUNet 3D is consistently among the best-performing DNNs across varying noise conditions, indicating superior adaptability to the complex filament structures in the actin network.

TABLE II:

Comparison of segmentation performance (Dice and IoU) of five deep learning architectures on Dataset-2 (actin network) under different noise conditions

Noise Level	Model	Metric	Background	Filament	Micro (Overall)	Macro (Class Avg.)
0.35	3D U-Net	Dice	0.991	0.618	0.991	0.804
		IoU	0.982	0.447	0.982	0.714
	Attention 3D U-Net	Dice	0.990	0.626	0.991	0.808
		IoU	0.981	0.456	0.981	0.718
	TransUNet 3D	Dice	0.991	0.673	0.991	0.832
		IoU	0.982	0.507	0.982	0.744
	UNETR	Dice	0.991	0.619	0.991	0.805
		IoU	0.981	0.448	0.982	0.715
	SegFormer3D	Dice	0.989	0.464	0.989	0.727
		IoU	0.977	0.302	0.978	0.640
0.50	3D U-Net	Dice	0.989	0.542	0.990	0.765
		IoU	0.979	0.371	0.979	0.675
	Attention 3D U-Net	Dice	0.989	0.576	0.990	0.783
		IoU	0.979	0.405	0.979	0.692
	TransUNet 3D	Dice	0.990	0.602	0.990	0.796
		IoU	0.979	0.430	0.980	0.705
	UNETR	Dice	0.990	0.602	0.990	0.796
		IoU	0.980	0.430	0.980	0.705
	SegFormer3D	Dice	0.990	0.573	0.990	0.782
		IoU	0.980	0.402	0.980	0.691
0.65	3D U-Net	Dice	0.988	0.540	0.988	0.764
		IoU	0.976	0.370	0.976	0.673
	Attention 3D U-Net	Dice	0.989	0.541	0.989	0.765
		IoU	0.978	0.370	0.978	0.674
	TransUNet 3D	Dice	0.988	0.534	0.989	0.761
		IoU	0.977	0.365	0.977	0.670
	UNETR	Dice	0.989	0.529	0.989	0.759
		IoU	0.978	0.359	0.978	0.669
	SegFormer3D	Dice	0.987	0.250	0.987	0.618
		IoU	0.973	0.143	0.974	0.558
0.80	3D U-Net	Dice	0.988	0.454	0.9880	0.721
		IoU	0.976	0.294	0.976	0.635
	Attention 3D U-Net	Dice	0.988	0.444	0.988	0.716
		IoU	0.975	0.285	0.976	0.630
	TransUNet 3D	Dice	0.984	0.536	0.984	0.760
		IoU	0.968	0.366	0.968	0.667
	UNETR	Dice	0.988	0.474	0.988	0.731
		IoU	0.976	0.310	0.976	0.643
	SegFormer3D	Dice	0.987	0.332	0.987	0.660
		IoU	0.974	0.199	0.974	0.587
0.95	3D U-Net	Dice	0.987	0.447	0.987	0.717
		IoU	0.974	0.288	0.974	0.631
	Attention 3D U-Net	Dice	0.983	0.463	0.984	0.723
		IoU	0.967	0.301	0.968	0.634
	TransUNet 3D	Dice	0.986	0.478	0.986	0.732
		IoU	0.973	0.314	0.973	0.643
	UNETR	Dice	0.986	0.487	0.986	0.737
		IoU	0.972	0.322	0.973	0.647
	SegFormer3D	Dice	0.986	0.291	0.986	0.639
		IoU	0.972	0.171	0.972	0.571

C. Overall Comparison and Observations

We observe that the segmentation performance of the deep learning architectures consistently decreases as the noise level increases, highlighting the challenge of filament detection under noisy conditions. Moreover, the results indicate that the performance of these architectures depends on both the filament arrangements and the noise level. Fig. 2 visualizes the comparative performance of the five semantic segmentation architectures in terms of the most important evaluation metric for this task, the filament Dice score. For Dataset-1, the performance of Attention 3D U-Net and 3D U-Net varies across the noise levels, indicating that their relative effectiveness depends on the noise conditions. Attention 3D U-Net achieves higher Dice scores at noise levels 0.4 and 0.8, whereas 3D U-Net performs better at 0.6 and 1.0. These results indicate that neither model demonstrates a consistent robustness trend across noise conditions. For Dataset-2, TransUNet 3D achieves the highest accuracy at both low (0.35) and high (0.80) noise levels, while UNETR shows strong performance under moderate to high noise (0.5 and 0.95). The Attention 3D U-Net achieved the highest Dice score at a noise level of 0.65. SegFormer3D consistently exhibits lower performance across both datasets, particularly for the filament class.

Fig. 2:

Filament Dice scores across varying noise levels for both datasets (* indicates the best-performing architecture)

VI. Summary and conclusion

We evaluated five deep learning architectures—3D U-Net, Attention 3D U-Net, UNETR, TransUNet 3D, and SegFormer3D—on two sets of simulated cryo-ET tomograms representing distinct filament arrangements under varying noise conditions. Our results indicate that model performance is highly dependent on filament arrangement and noise levels, with no single architecture demonstrating consistent superiority across all datasets and conditions. For the semi-regular actin bundles (Dataset-1), purely convolutional architectures such as 3D U-Net and Attention 3D U-Net achieved superior segmentation accuracy. In contrast, for the more complex actin network (Dataset-2), hybrid convolutional-transformer architectures UNETR and TransUNet 3D performed better at low-to-moderate noise levels and showed greater robustness under higher noise.

Overall, this benchmark provides practical guidance for selecting semantic segmentation architectures for cryo-ET–based actin filament analysis and informs the development of hybrid architectures optimized for structures with diverse arrangements and noise conditions. Although here we evaluated all the DL architectures primarily using their default hyperparameter settings, fine-tuning may be necessary to identify the most suitable model for specific contexts. Future work will focus on fine-tuning, multi-scale feature learning, and hybrid designs to further improve segmentation accuracy under challenging conditions.

Fig. 1:

Simulated tomograms of actin filaments. (A) Dataset 1: Simulated tomograms of semi-regular actin filaments representing the tapered regions of stereocilia under varying noise levels (top: model, bottom: corresponding tomograms). (B) Dataset 2: Simulated tomograms of the actin network of filopodia under varying noise levels (top: model, bottom: corresponding tomograms).

Fig. 3:

(A) Dataset-1 (noise level 0.6) showing predictions by the best-performing 3D U-Net: top left, density map; top right, ground truth overlaid on density map; bottom left, prediction overlaid on density map; bottom right, prediction overlaid on ground truth. (B) Dataset-2 (noise level 0.5) showing predictions by TransU-Net 3D: top left, density map; top right, ground truth overlaid on density map; bottom left, prediction overlaid on density map; bottom right, prediction overlaid on ground truth.