Abstract
Purpose
Filopodia are thin, finger-like extensions that project from the surface of cells. Present in various cell types, analyzing these structures can yield significant insights into cellular behavior and function. However, a manual analysis is impractical, resulting in a bottleneck in processing large-scale datasets.
Methods
This paper introduces FiloAnalyzer, an open-source toolbox for automatically analyzing cell filopodia. The toolbox is powered with a user-friendly GUI, allowing the segmentation of batches of microscopy images and extracting a set of quantitative parameters, incorporating both a deep learning-based and an image-processing method. As data and annotation scarcity are typical of this domain, we compare popular ImageNet supervised to self-supervised pre-training, adopting state-of-the-art contrastive, self-distillation, and generative approaches based on diffusion models in an attempt to maximize segmentation performance in a low-data regime.
Results
We validate our toolbox on a dataset of annotated fluorescent images of neural crest cells, benchmarking our results with three popular filopodia segmentation toolboxes. Furthermore, we show how self-supervision can be a promising tool to deal with the limited availability of annotated data in this domain, obtaining a significant improvement over ImageNet supervised pre-training, with as few as ten annotated images.
Conclusion
Our experiments show that the deep learning pipeline outperforms image processing-based available alternatives. We release both an annotated dataset and the toolbox as open-source to foster further research on this topic. The toolbox code and dataset are available at https://github.com/Malga-Vision/FiloAnalyzerToolbox.
Keywords: Filopodia segmentation and analysis, Deep learning, Image processing, Transfer learning
Introduction
Filopodia are thin, actin-rich cell membrane protrusions found in virtually all cell types [1, 2]. They act as sensory and traction structures that probe the environment, adhere to surfaces and pull the cell forward during migration, therefore playing fundamental roles in wound healing and cell signaling [3]. Because of their prominent role in cell motility, they have been associated with invasive cancers [4]. Furthermore, filopodia are also involved in Wnt signaling [5]. Wnt proteins, which are lipid-modified morphogens, are crucial for signaling, embryonic development, and cancer progression [6–8]. Although Wnt proteins largely accumulate in the endoplasmic reticulum (ER) when overexpressed [9], co-expression of Wntless (WLS), a Wnt transporter along with WNT1-GFP, causes a redistribution of WNT1-GFP from the ER to the cell surface and filopodia [5, 10–12]. Additionally, WLS augments the signaling activity of WNT1 [13]. The presence and importance of Wnts and WLS in filopodia have been demonstrated in various model organisms, including chick, zebrafish, and drosophila [5, 14–16]. Consequently, analyzing these cellular extensions can provide important information about cell state, behavior, and, more specifically, Wnt signaling.
However, manual analysis is time-consuming and resource-intensive, hindering the investigation of large-scale datasets, pushing toward the development of automatic solutions. Rather than exclusive of filopodia, such need for automatic analysis is common to general cell protrusions detection tasks, with several methods released in this domain of application. In [17], a deep learning approach is proposed for measuring hypocotyls lengths in plantlet. Specifically, a Unet is trained to segment hypocotyls, with the prediction masks skeletonized to get their lenghts. In [18], a 3D segmentation model derived from the nnUnet [19] is employed for segmenting the plasma membrane surface, further refined to detect microvilli protrusions. In [20] the authors introduce u-shape3D, a method based on image processing to segment cell surface, further decomposed in convex patches, and a supervised machine learning framework based on Support Vector Machine (SVM) on top of geometric patch features to detect cell motifs, including blebs, lamellipodia and filopodia. In [21], an annotation-free method based on edge detection and contour quantification is proposed for microvilli segmentation, to face the limited availability of data typical of the domain of application. Focusing more specifically on cell filopodia, several automated methods of analysis have been developed in the last years [22–25]. Nilugar and colleagues introduce a MATLAB [26] toolbox based on image processing, with automatic thresholding for cell body segmentation, and a series of morphological operations to identify filopodia, providing measurements of their length and number [22]. Various image analysis tools using ImageJ [27] as an image analysis framework have also been released, with features ranging from providing cell body, filopodia base, corpus, and tips segmentation to their estimated number [23, 24]. Modern deep learning based methods have also been recently employed. For example, the work in [25] uses the ZeroCostDL4Mic [28] platform to train an Unet [29] architecture in a semantic segmentation framework including three classes: background, cell body, and filopodia tips. Despite the availability of several works for protrusion segmentation, the problem of addressing specific challenges for the automated analysis of cell filopodia images is yet underexplored, especially in terms of how to train robust and accurate deep learning methods with limited annotated data in this context, and the community has relied on mostly all-purpose tools built for general image analysis tasks.
In this work, we present FiloAnalyzer, an open-source toolbox for 2D filopodia segmentation and analysis, incorporating both an image-processing and a deep learning approach. FiloAnalyzer is developed in Python and offers a friendly Graphical User Interface (GUI) (see Fig. 1), allowing even non-expert users to automatically analyze large batches of images containing cell filopodia. Differently from [22–24], we use a deep learning framework that allows accurate filopodia mask prediction and where we test multiple Fully Convolutional Neural Networks (FCN) and pre-trained backbones for the filopodia segmentation downstream task. Specifically, to address the complexity of the task at hand, we employ different pre-training strategies before fine-tuning the obtained model on the downstream task. First, similarly to [28], we make use of a transfer learning framework, starting from deep neural networks pre-trained on ImageNet [30] with supervision. However, a domain shift between ImageNet and filopodia fluorescence images exists, potentially impacting the model performance on the segmentation downstream task. To overcome this issue, we explore in-domain self-supervised protocols, exploiting the same filopodia image dataset with no annotations as a source dataset and fine-tuning the resulting model on the filopodia segmentation task.
Fig. 1.
Toolbox GUI showing the intermediate steps of FiloAnalyzer’s image processing pipeline: (Top-left) Input image; (Top-right) Segmented cell image; (Bottom-left) Segmented Image after erosion and dilation; (Bottom-Right). Final result after refinement, with identified filopodia. See Sect. “FiloAnalyzer’s image processing method” for more details on each step
Specifically, in our experiments, we evaluate four well-established self-supervised methods, SimCLR [31], SimCLRv2 [32], MoCO [33], and DINO [34], plus one state-of-the-art self-supervised pre-training method with a generative pretext task, based on diffusion models [35]. Finally, the resulting model is used for the filopodia segmentation task and fine-tuned on the annotated filopodia dataset. Exploiting self-supervised pre-training, our results confirm it is possible to train an accurate and robust model, even with a very limited availability of annotated images. It is worth mentioning that, at the time of writing this paper, no open-source annotated benchmark datasets for filopodia segmentation are available, potentially hindering the adoption of deep learning frameworks for filopodia segmentation. To fill this gap, this work contributes and releases a novel dataset, Filo-SEG, containing 245 fluorescent images of chicken embryonic neural crest cells with expert-annotated filopodia. To our knowledge, this is the first publicly released fully annotated dataset for filopodia segmentation that can provide a benchmark for the scientific community while fostering further research in this context. To summarize, our main contributions include:
An open-source dataset of annotated fluorescent filopodia images.
A deep learning method based on transfer learning, self-supervision adopting state-of-the-art methods with contrastive, self-distillation and generative pretext tasks, capable of providing accurate and robust results with few annotated data.
An open-source toolbox for filopodia analysis, implementing a deep learning method for accurate filopodia segmentation, providing prediction masks and number of detected filopodia. We benchmark FiloAnalyzer with state-of-the-art filopodia segmentation toolboxes on our Filo-SEG dataset, obtaining a significant improvement in terms of filopodia precision and recall.
FiloAnalyzer and the Filo-SEG dataset are available at https://github.com/Malga-Vision/FiloAnalyzerToolbox.git.
The remainder of the paper is organized as follows: in Sect. “Experiments”, we describe the datasets, provide experiment details and evaluation metrics, and report the obtained results. Then, we discuss our results in Sect. “Discussion” and provide conclusions in Sect. Conclusions. Finally, in Sect. “Methods”, we describe our image-processing and deep learning pipelines, the filopodia counting approach, and FiloAnalyzer’s GUI.
Experiments
This section describes our dataset, implementation details, and training protocols. Then, we report the obtained results.
Dataset
To our knowledge, no open-source annotated dataset exists for filopodia segmentation. Consequently, we acquired and used two in-house datasets to train and validate the proposed approach, which we refer to as the Filopodia Segmentation (Filo-SEG) and the WNT1 dataset. All the technical details on data acquisition, including materials, cell culture procedure, and annotations, are available in the Supplementary Materials.
Filopodia SEGmentation (Filo-SEG) dataset
The Filo-SEG dataset contains 245 annotated fluorescent images of HEK 293T, NIH 3T3, Hamburger & Hamilton (HH) 13-15 chicken embryonic neural crest cells, divided into training (200) and test (45) images. The images were collected using a Zeiss LSM 710 confocal with either a 63x dipping objective or a 63x Plan Apochromat oil objective at room temperature. The dataset presents both single-cell images, in which the filopodia are relatively easy to detect, as well as images with multiple cells, in which the filopodia are harder to segment and "disentangle". Images were annotated using Adobe Photoshop, tracing each filopodium mask with a brush size of 4 pixels and white color. However, FiloAnalyzer accepts annotations as visual masks which can be obtained through any suitable image analysis software, including ImageJ. See Fig. 2 for sample images.
Fig. 2.
Sample images of the Filo-SEG dataset (left) and corresponding annotation masks (right)
WNT1 dataset
The second dataset, which has been introduced in [5], contains a total of 31 fluorescent images from cells transfected with WNT1-GFP and mKate-CAAX 2. The number of filopodia in each image in the dataset has been manually counted by our experts. We use this dataset for method validation, that is, to evaluate FiloAnalyzer’s accuracy in counting the number of filopodia, comparing its predictions with the available ground truth information. See Fig. 3 for sample images.
Fig. 3.
Sample images of the WNT1 dataset
Dataset details and preprocessing
We apply the following preprocessing pipeline for the filopodia segmentation downstream task, both for ImageNet and self-supervised pre-training. We resize the filopodia images and the corresponding segmentation mask to 
pixels, then we normalize the images to the interval [0, 1], and finally standardize them using the ImageNet mean ([0.485, 0.456, 0.406]) and standard deviation ([0.229, 0.224, 0.225]). During the training, we apply batch-level data augmentation:
When training the segmentation model, we apply random data augmentation, including horizontal and vertical flips, rotation by random angles up to
, and elastic transformations to introduce random deformations.During the self-supervised pre-training, each method applies its own set of random data augmentation. SimCLR and SimCLRv2 apply random cropping and resizing, horizontal flips, gray-scale conversion, Gaussian blur, and color jitter that change the image’s brightness, contrast, saturation, and hue. MoCo adopts the same augmentations except for the Gaussian blur. DINO considers 3 sets of data augmentations, similar to the ones present in SimCLR, to learn global features at high magnification and local features at low magnification. The diffusion model uses shifting, scaling, rotation, resizing, horizontal and vertical flips, random gamma adjustment, Gaussian blur, and standard normalization. These augmentations give the model rich variations of the image and force it to learn invariant and robust feature representations.
Experiment details and training protocol
We begin by holding out 45 randomly selected images for the test set (about 
of the dataset). To train the model, we employ a 10-fold Cross-Validation strategy to maximize the use of the available data while minimizing the effect of random data splits to ensure a robust evaluation. The experiments include four segmentation models, resulting from two encoders, ResNet152 and ResNest200e, incorporated into two Fully Convolutional Neural networks, UNet and UNet++.
When fine-tuning the pre-trained models on the downstream task, we employ 200 epochs with a batch size of 32, the Adam optimizer with an initial learning rate of
that we control using the Exponential Learning Rate Scheduler. We monitor the Dice loss at each epoch, saving the model with the lowest loss value. All the encoders and FCN implementations are provided by the segmentation_model_pytorch[36] library, which provides efficient implementations of state-of-the-art segmentation models.For SimCLR, MoCo, and DINO, we use the implementation from the GitHub repository [37]. Self-supervised pre-training requires no labels, which allows us to benefit from the entire Filo-SEG images. We run the pre-training for 200 epochs, with a fixed learning rate of
with the Adam optimizer. At this stage, we only track the training loss, saving the model with the lowest value. The standard self-supervision methods require careful choice of the projection dimension (the size of the features space where we transform the images) and the temperature (a scaling parameter of the contrastive loss for the separation between the positive and negative pairs of images). We run a grid search over a range of values for the projection dimensions ([32, 64]) and the temperature ([0.4 0.3 0.2 0.1] for SimCLR and SimCLRv2, and [0.04 0.03 0.02 0.01] for MoCo), selecting the combination that yields the best performance. We use the default values of DINO as they produce the best results.Furthermore, building upon traditional self-supervised approaches, we experiment with a denoising diffusion probabilistic model (DDPM) pre-trained in a self-supervised way, adhering to the experimental framework outlined in [35]. Unlike the previously mentioned approaches, DDPMs learn representations by modeling the data distribution itself through a denoising objective, progressively generating clean images from noise. The learned feature representations are then transferred to the downstream segmentation task by fine-tuning the denoising network on labeled data. This paradigm has recently shown strong robustness in low-data regimes [35], making it particularly suitable for filopodia segmentation, where the availability of annotations is limited by costs in terms of time and resources. For the diffusion-based pre-training, we adopt a UNet-based denoising architecture following standard DDPM design principles. Specifically, we employ multiple resolution levels with channel multipliers
and four residual blocks per level. Self-conditioning is enabled to improve denoising stability, and self-attention mechanisms are applied at a spatial resolution of 
pixels. The diffusion process is defined over 1000 timesteps using a linear noise schedule with 
values ranging from 
to 0.02. The model is trained using an 
denoising loss, and an exponential moving average (EMA) of the model parameters is maintained during training to improve convergence and sampling stability. For more details on DDPM, please refer to [38].
Evaluation metrics
We exploit four different evaluation metrics to evaluate our results: the Dice Score (DSC), Precision, Recall, and F1 Score [39]. All these metrics are ideal for binary classification and have values ranging between 0 and 1, where higher values indicate better performance.
Dice score
The Dice Score (DSC) (Eq. 1) is a similarity measure between two images containing the segmentation areas, where one is the ground truth (GT) and the other is the model’s prediction output (Pred). It is a ratio between two quantities: (i) the first quantity, 
is double the intersection between Pred and GT, which is the number of pixels in common between the two segmentation areas; (ii) the second quantity, 
, is the total number of pixels present in both areas.
 |
1 |
Precision
The Precision (Eq. 2) is a quantification of how many of the pixels predicted as positive are indeed positive, with respect to all positive predictions. The True Positives are the foreground (or the object) pixels that were correctly predicted as foreground. The False Positives are the background pixels that were misclassified. The sum of TP and FP is the total number of pixels that were predicted to belong to the object. High Precision indicates the model is excellent at predicting the correct pixels (low False Positives).
 |
2 |
Recall
The Recall(Eq. 3) is a quantification of how many pixels were predicted as positive out of all the positive pixels in the image. The False Negatives are the foreground pixels that were predicted as background. The sum of TP and FN is the total number of pixels belonging to the object we want to segment. High Recall means that the model has identified all the positive pixels correctly (low False Negatives).
 |
3 |
F1 score
The F1 Score(Eq. 4) is the harmonic mean between Precision and Recall. It has the effect of emphasizing the smaller of the two values, meaning that a low F1 Score implies that either the Precision or the Recall is low, pointing out that the model has high False Positives or high False Negatives.
 |
4 |
Results
In this section, we first report the results corresponding to the usage of different pre-training paradigms for the segmentation encoder, along with the multiple combinations of encoders and architectures for our deep learning filopodia segmentation approach on the Filo-SEG dataset. Then, we show a comparison in terms of performance with state-of-the-art open-source toolboxes on the same dataset. Finally, we provide results for an experiment on the WNT-1 dataset to validate the FiloAnalyzer’s deep learning model in terms of the number of predicted filopodia.
Comparative study on pre-training, backbones and architectures
Table 1 summarizes the comparative study’s results of the various pre-training paradigms and the encoder-architecture combinations for our deep learning approach on the Filo-SEG dataset. Results obtained using generative diffusion-based pre-training (DDPM) are not included in this table, as they rely on a fundamentally different architecture and training objective, and are therefore evaluated separately. For each experiment, we report the mean ± standard deviation on the validation set, obtained with the 10-fold cross-validation procedure. Our evaluation is based on the Dice Score, though we also report other metrics such as Precision, Recall, and F1 Score.
Table 1.
Performance comparison of different training paradigms (Supervised, SimCLR, SimCLRv2, MoCo, and DINO) for filopodia segmentation on the 10-fold validation sets
| Decoder | Encoder | Training paradigm | Dice score | Precision | Recall | F1 score |
|---|---|---|---|---|---|---|
| UNet++ | ResNet152 | ImageNet weights | 0.691 ± 0.01 | 0.7 ± 0.016 | 0.73 ± 0.014 | 0.695 ± 0.006 |
ImageNet SimCLR |
0.695 ± 0.006 | 0.712 ± 0.012 | 0.738 ± 0.013 | 0.708 ± 0.005 | ||
ImageNet SimCLRv2 |
0.69 ± 0.005 | 0.711 ± 0.017 | 0.733 ± 0.014 | 0.702 ± 0.007 | ||
ImageNet MoCo |
0.694 ± 0.011 | 0.712 ± 0.01 | 0.738 ± 0.018 | 0.707 ± 0.011 | ||
ImageNet DINO |
0.687 ± 0.009 | 0.7 ± ± 0.022 | 0.736 ± 0.018 | 0.7 ± 0.008 | ||
| ResNest200e | ImageNet weights | 0.701 ± 0.019 | 0.703 ± 0.032 | 0.762 ± 0.028 | 0.715 ± 0.02 | |
ImageNet SimCLR |
0.714 ± 0.009 | 0.734 ± 0.015 | 0.752 ± 0.01 | 0.729 ± 0.01 | ||
ImageNet SimCLRv2 |
0.715 ± 0.011 | 0.739 ± 0.015 | 0.748 ± 0.018 | 0.728 ± 0.012 | ||
ImageNet MoCo |
0.718 ± 0.008 | 0.737 ± 0.012 | 0.755 ± 0.016 | 0.732 ± 0.008 | ||
ImageNet DINO |
0.718 ± 0.007 | 0.743 ± 0.015 | 0.75 ± 0.014 | 0.732 ± 0.006 | ||
| Random weights | 0.666 ± 0.019 | 0.675 ± 0.024 | 0.706 ± 0.02 | 0.677 ± 0.019 | ||
Random weights  DINO |
0.722 ± 0.02 | 0.715 ± 0.026 | 0.773 ± 0.03 | 0.731 ± 0.02 | ||
| ResNet18 | ImageNet weights | 0.683 ± 0.017 | 0.691 ± 0.026 | 0.727 ± 0.023 | 0.695 ± 0.017 | |
| Random weights | 0.639 ± 0.014 | 0.667 ± 0.022 | 0.667 ± 0.025 | 0.65 ± 0.013 | ||
| UNet | ResNet152 | ImageNet weights | 0.679 ± 0.013 | 0.69 ± 0.014 | 0.71 ± 0.011 | 0.676 ± 0.005 |
ImageNet SimCLR |
0.683 ± 0.01 | 0.697 ± 0.021 | 0.731 ± 0.012 | 0.698 ± 0.009 | ||
ImageNet SimCLRv2 |
0.689 ± 0.006 | 0.697 ± 0.012 | 0.698 ± 0.009 | 0.703 ± 0.006 | ||
ImageNet MoCo |
0.681 ± 0.007 | 0.689 ± 0.009 | 0.736 ± 0.011 | 0.695 ± 0.006 | ||
ImageNet DINO |
0.686 ± 0.005 | 0.696 ± 0.013 | 0.736 ± 0.015 | 0.7 ± 0.005 | ||
| ResNest200e | ImageNet weights | 0.701 ± 0.007 | 0.732 ± 0.009 | 0.7 ± 0.005 | 0.72 ± 0.006 | |
ImageNet SimCLR |
0.712 ± 0.009 | 0.731 ± 0.014 | 0.749 ± 0.014 | 0.727 ± 0.008 | ||
ImageNet SimCLRv2 |
0.708 ± 0.016 | 0.734 ± 0.015 | 0.74 ± 0.013 | 0.723 ± 0.014 | ||
ImageNet MoCo |
0.709 ± 0.009 | 0.73 ± 0.006 | 0.746 ± 0.012 | 0.72 ± 0.006 | ||
ImageNet DINO |
0.708 ± 0.008 | 0.73 ± 0.013 | 0.745 ± 0.013 | 0.723 ± 0.005 |
The arrow indicates that the encoder is pre-trained on ImageNet before training in a self-supervised paradigm. Metrics include Dice Score, Precision, Recall, and F1 Score. The best-performing results for each metric are highlighted in bold
The aim of such experiments is to select the best architecture, backbone and self-supervised learning (SSL) pre-training protocol, according to the average results in terms of Dice Score on the validation sets. We explore two commonly used Fully Convolutional Network architectures in the semantic segmentation of biomedical images, the UNet[29] and UNet++[40]. UNet was originally proposed for biomedical images and quickly gained popularity for its simple architecture and high performance. The UNet has a symmetric architecture, with the same number of layers in both the encoder and decoder. Their key innovation is the concept of skip-connections: each convolutional layer of the decoder is concatenated with the feature map from its corresponding layer in the encoder before applying the convolution operation. UNet++ is an evolution of UNet with dense and more complex skip-connections. Regarding the backbones, we consider two deep Convolutional Neural network backbones, ResNet152[41] and ResNest200e[42]. The ResNet has proven its robustness over a wide range of applications, mainly due to the Residual Blocks, a mechanism that allows a ResNet model to grow deeper, capturing more features while training efficiently. The residual block design feature has spawned other derived architectures like ResNest, which we also evaluate in our experiments.
We implement a Transfer Learning approach, a commonly used strategy to deal with the limited availability of annotated data [43]. We exploit two pre-training paradigms: ImageNet supervised pre-training and SSL pre-training. In the first transfer learning approach, we start with an encoder pre-trained on ImageNet and fine-tune the entire FCN architecture on the target dataset. In Table 1, we refer to this setup with ImageNet weights. The second transfer learning approach involves two steps: (i) using a self-supervised approach, we pre-train the encoder on the Filo-SEG dataset with no annotations, where the model learns a meaningful representation of the filopodia images; (ii) we fine-tune the pre-trained encoder with the segmentation architecture on the downstream task of filopodia segmentation. In Table 1, we refer to this setup ImageNet 
SSL method, where the arrow signifies that the encoder was initially pre-trained with ImageNet and then used for the SSL setup. For the self-supervised pre-training, we explore four methods: SimCLR, SimCLRv2, MoCo, and DINO. Simple framework for Contrastive Learning of Visual Representation (SimCLR[31]) and (SimCLRv2[32]) are based on Contrastive Learning, where the model learns to maximize the similarity between positive pairs of images and minimize the similarity between negative pairs; a positive pair is formed by applying two sets of random data augmentation to the same image. Momentum Contract (MoCo)[33] is also based on Contrastive Learning, adding the Momentum mechanism that acts as a memory bank to store previous negative pairs associated with an image, eliminating the need for large batch sizes. Self-Distillation with No Labels (DINO)[34] takes a different approach to the previous SSL approaches, adopting Knowledge Distillation to train a student network to mimic a teacher network: each network receives a transformed version of the same image at different scales (to learn global and local features), and the student learns the dataset representation guided by the teacher.
The ResNest200e consistently beats the ResNet152 in terms of Dice Score across all the training paradigms. This result hints that its architecture can better represent the filopodia filaments. The ResNest200e-UNet++ segmentation architecture pre-trained with self-supervised learning achieves the best Dice Score. DINO pre-training reaches 
, marginally surpassing MoCo (
). From a decoder point of view, the added complexity of UNet++ is better at capturing and segmenting the thin structure of the filopodia. Pre-training the encoders using self-supervised learning surpasses the ImageNet pre-training in almost all the experiments, showing the advantage of the in-domain data representation in recognizing the fine details of the filopodia structure.
We perform an additional experiment considering a shallower encoder, that is a ResNet18, to provide insights on the dependency of the final performance on the backbone’s depth. As reported in Table 1, for both randomly initialized and ImageNet pre-trained models, we notice a decreasing for all the evaluation metrics.
Moreover, to further evaluate the impact of SSL pre-training on the performance of our models, we provide a baseline corresponding to our best model (Unet++ and ResNest200e), trained with plain supervision starting from randomly initialized weights. The results reported in Table 1, confirm the benefit of SSL pre-training, with an average improvement of 
in Dice Score.
Additionally, we perform an ablation study where our SSL pre-training framework is employed on a randomly initialized backbone, instead of an ImageNet pre-trained one. The results summarized in Table 1 are comparable with the ones obtained with ImageNet pre-training, with a slightly higher precision corresponding to 
.
Finally, we conduct an ablation study to measure the impact of annotation masks thickness on the final performance. Specifically, employing the skeletonize method in Scikit-image [44], we reduce the annotation masks width and re-train our best model with the 10-folds cross-validation procedure, obtaining an average Dice Score of 
, empirically showing how a thicker mask is beneficial, as it increases the number of pixels belonging to the filopodia class of interest.
Analysis of performance trends in low-data regimes
Expanding on the insights from our comparative study, we now delve deeper into the implications of model performance across different data availability conditions. In a realistic scenario, with different acquisition systems or fluorescence labels, potential domain shifts can arise, making it necessary to fine-tune the FiloAnalyze provided model for maintaining good performance. However, obtaining high-quality annotations is expensive and subject to the annotator’s experience, and as such, we are interested in evaluating pre-training protocols in the low-data regime. Inspired by the results presented in [35], we compare an ImageNet supervised pre-training with a self-supervised method with a diffusion generative pretext task. Specifically, the pretext task consists of training a diffusion model to generate filopodia fluorescence images, exploiting the available dataset but discarding the annotations. The corresponding Unet is then fine-tuned on the filopodia segmentation downstream task. Table 2 presents the results for varying amounts of labeled samples on the test set, highlighting the trade-offs between these paradigms. From the full dataset results, it is evident that the ImageNet-supervised model achieves superior performance across all metrics, with a Dice Score of 0.701 and an F1 Score of 0.715. However, as the number of available labeled images decreases, the self-supervised DDPM approach demonstrates increased robustness. Notably, for 50 labeled samples, DDPM outperforms the supervised model in Dice Score (0.498 vs. 0.433), though it lags in the other metrics.
Table 2.
Comparison of test segmentation performance between the best ImageNet-supervised model and self-supervised DDPM pre-training across different amounts of labeled training samples
| Training labels | Training paradigm | Dice score | Precision | Recall | F1 score |
|---|---|---|---|---|---|
| full dataset | Best Imagenet | 0.703 ± 0.002 | 0.740 ± 0.011 | 0.757 ± 0.013 | 0.737 ± 0.003 |
| DDPM | 0.647 ± 0.034 | 0.648 ± 0.034 | 0.705 ± 0.023 | 0.627 ± 0.043 | |
| 50 | Best Imagenet | 0.433 ± 0.056 | 0.687 ± 0.010 | 0.740 ± 0.017 | 0.678 ± 0.024 |
| DDPM | 0.498 ± 0.037 | 0.506 ± ± 0.033 | 0.601 ± ± ± 0.022 | 0.472 ± 0.048 | |
| 25 | Best Imagenet | 0.173 ± 0.052 | 0.581 ± 0.069 | 0.831 ± 0.051 | 0.470 ± 0.088 |
| DDPM | 0.443 ± 0.047 | 0.448 ± 0.049 | 0.548 ± 0.043 | 0.437 ± 0.072 | |
| 10 | Best Imagenet | 0.156 ± 0.051 | 0.517 ± 0.071 | 0.792 ± 0.047 | 0.428 ± 0.085 |
| DDPM | 0.311 ± 0.064 | 0.423 ± 0.223 | 0.461 ± 0.036 | 0.301 ± 0.104 |
The evaluation considers Dice Score, Precision, Recall, and F1 Score. DDPM exhibits higher robustness in limited data scenarios, particularly in extremely low-data settings
The trend becomes even more pronounced in extreme low-data regimes. With only 25 and 10 labeled images, DDPM consistently achieves higher Dice Scores and maintains more stable recall values. For instance, at 10 labels, DDPM achieves a Dice Score of 0.311 compared to 0.156 for the supervised model, indicating better segmentation reliability under severe annotation constraints.
These findings suggest that while supervised models dominate in large-data settings, self-supervised diffusion models provide a compelling alternative when labeled data is scarce. This aligns with the broader trend in deep learning where self-supervised approaches increasingly mitigate reliance on extensive annotation, particularly in medical and low-resource imaging scenarios.
Benchmarking FiloAnalyzer with existing open-source toolboxes
In our next experiment, we benchmark FiloAnalyzer’s deep learning approach with three popular open-source toolboxes for filopodia analysis, namely FiloDetect [22], FiloQuant [24], and Filopodyan [23]. Furthermore, we train and evaluate a model based on a custom implementation of FiloVision’s Unet [25] and an image-processing-based approach for filopodia segmentation incorporated in FiloAnalyzer to provide a computationally light alternative to the deep learning method. Table 3 summarizes the results obtained on the Filo-SEG test set. For a fair comparison, results with Filodetect, FiloQuant, Filopodyan, and our image processing approach are obtained by manually tuning the methods’ parameters for each image, aiming at maximizing the prediction accuracy. Our deep learning method largely outperforms all the remaining approaches, showing an improvement of 
in terms of Dice Score over the best-performing among the open-source available toolboxes (FiloQuant) and an improvement of 
and 
for precision and recall, respectively. Moreover, the standard deviation among test images is much lower than the benchmarking methods, indicating that the proposed approach is robust with respect to the test images. Finally, our image processing approach obtains an average improvement in terms of Dice, and Recall, but a high standard deviation and in line with the existing methods.
Table 3.
FiloAnalyzer performance with respect to open-source popular filopodia analysis toolboxes on the Filo-SEG test set (45 images)
| Tool | Dice | Precision | Recall | F1 Score |
|---|---|---|---|---|
| FiloDetect | 0.250 ± 0.177 | 0.497 ± 0.299 | 0.253 ± 0.210 | 0.256 ± 0.175 |
| FiloQuant | 0.279 ± 0.194 | 0.537 ± 0.274 | 0.257 ± 0.198 | 0.279 ± 0.194 |
| Filopodyan | 0.226 ± 0.163 | 0.352 ± 0.301 | 0.339 ± 0.259 | 0.237 ± 0.159 |
| Unet from [25] | 0.357 ± 0.195 | 0.699 ± 0.235 | 0.285 ± 0.162 | 0.357 ± 0.195 |
| FiloAnalyzer (Image processing) | 0.345 ± 0.211 | 0.479 ± 0.289 | 0.420 ± 0.262 | 0.353 ± 0.206 |
| FiloAnalyzer (Deep learning) | 0.715 ± 0.107 | 0.732 ± 0.085 | 0.750 ± 0.161 | 0.730 ± 0.107 |
Validation of FiloAnalyzer on the WNT1 dataset
We employ FiloAnalyzer’s deep learning approach for analyzing the WNT1 dataset (see Sect. “Experiments”). Specifically, we exploited this dataset to evaluate FiloAnalyzer’s performances regarding the prediction of the number of filopodia. We employ the best Unet++ with ResNest200e on the FiloSEG 10 folds experiment to predict the number of filopodia per image in the WNT1 dataset, comparing it with the ground truth manually determined count. Figure 4 summarizes the obtained results. FiloAnalyzer’s predictions strongly correlate with the manual annotations (
.
Fig. 4.
Examples of output masks from the Filo-SEG test set between FiloAnalyzer and existing open-source toolboxes employing image processing-based methods
empirically validating the implemented counting filopodia approach and supporting the potential of FiloAnalyzer for processing large-scale cell filopodia image datasets. Table 4 reports the predicted and the ground truth average number of filopodia.
Table 4.
Comparison between average filopodia counts, obtained by a human expert annotator (Manual count) and FiloAnalyzer
| Manual count | FiloAnalyzer |
|---|---|
| 21.5 | 21.4 |
Discussion
A manual analysis of cell filopodia is impractical, pushing toward the need for automatic strategies, including deep learning approaches. However, deep learning methods are notoriously data-hungry, and a direct application in the context of filopodia analysis is hindered by the lack of annotated datasets. In this work, we address these gaps with a twofold contribution, including the release of an open-source fully annotated dataset for filopodia segmentation, and the proposal of SSL as an effective pre-training strategy for obtaining robust segmentation models, capable of providing high accuracy for the task at hand. Our results confirm how using an FCN with an SSL pre-trained backbone provides improvements over conventional image processing methods, such as [22–24], with higher precision, recall and dice score in filopodia detection, even if trained on a few hundred annotated images. The benefit of SSL for the downstream task is further highlighted by the improvements over the Unet model employed in a recent filopodia segmentation deep learning-based method [25]. To better frame the difference in performance with conventional image processing methods, in Fig. 5, we provide four examples of output masks for our deep learning- and image processing-based methods, and the benchmark open-source toolboxes that we consider in this work. Importantly, as noticeable in the first row, despite the deep learning approach still providing the best performance on average, in case of well-contrasted images with few cells, the conventional methods are capable of providing acceptable performance. On the other hand, when multiple cells or more noise is present, their performance degrades significantly, while the deep learning method remains robust, as reflected by the numeric results reported in Table 3.
Fig. 5.
Number of predicted filopodia versus ground truth in the WNT1 dataset
Our findings are coherent with recent works on protrusions segmentation, confirming the high potential of deep learning based methods in this context [17, 18]. Furthermore, on top of the investigated conventional SSL strategies, our experiments with generative pre-training employing diffusion models provide a promising direction for the application of deep learning within few-shot regimes, in case of very limited annotated data, as common in the biological domain of application, reaching a dice score higher than benchmark filopodia segmentation image processing methods, with as few as 10 annotated images (see Table 2).
In principle, we expect our approach to be extendable to other cell protrusions, including microvilli, lamellipodia, blebs and neural dendrites. However, our trained models should not be interpreted as universal cell protrusion detectors. Their performance is hindered by the intrinsic large variability affecting fluorescent images, with potential domain shifts induced by several factors, including imaging hardware, fluorescent labels and acquisition conditions. Therefore, we recommend employing FiloAnalyzer’s code to either fine-tune the available model (or re-train a model with the implemented SSL framework) on an annotated collection of protrusions of interest. Rather than a universal plug-and-play product for filopodia or general protrusions, our aim is primarily to provide a dynamic and modular framework that can be expanded with more annotated data for improved generality and performance.
Conclusions
Filopodia are thin projections from the cell surface that play various and crucial roles in the cell’s interaction with its environment. Analyzing such cellular structures may provide fundamental information on cell state. However, manual analysis and counting are impractical, especially for large-scale datasets.
In this paper, we present a toolbox for automatic segmentation and analysis of cell filopodia, which we named FiloAnalyzer. It is coupled with a user-friendly GUI, allowing the usage of our deep learning-based approach for accurate filopodia segmentation. Additionally, we incorporated an image-processing pipeline in FiloAnalyzer, to be used as a computationally lighter alternative for fast inference. In this work, we also introduce Filo-SEG, the first open-source filopodia dataset, including 245 fluorescent images annotated by experts. To set a benchmark on our dataset, we pre-train a ResNest200e on our dataset using the self-supervised method DINO before using it as the encoder for a UNet++ during the downstream task of filopodia segmentation. This approach outperforms our image processing pipeline and the other state-of-the-art approaches.
FiloAnalyzer segments and counts all the filopodia present in the input image, further allowing the user to limit the analysis to a manually selected ROI. Future work will include nuclei detection to associate filopodia with specific cells. Furthermore, the model ends up detecting filopodia missed by the experts, suggesting that FiloAnalyzer could be incorporated into an active learning framework, assisting users in annotating more samples that could be used to fine-tune the model for further improving the segmentation performance. The toolbox code and dataset are available at https://github.com/Malga-Vision/FiloAnalyzerToolbox.
Methods
In this section, we provide details on the two methods embedded in FiloAnalyzer. Specifically, in Sect. “Deep learning pipeline” we describe the deep learning approach employing the SSL framework, while in Sect. “FiloAnalyzer’s image processing method” we detail the image processing method, included as a computationally lighter alternative when limited computational resources prevent the usage of deep learning.
Deep learning pipeline
We frame filopodia detection as a binary segmentation problem, training a Fully Convolutional Neural Network (FCN) [45] to obtain prediction masks. Figure 6 schematically represents the proposed approach. The filopodia are thin filaments that constitute a small portion of the image, which challenges any segmentation pipeline. We exploit an FCN for filopodia segmentation, with a deep architecture to learn global and local context to distinguish the filopodia from the cell body. Specifically, we implement a custom FCN that uses a UNet++ architecture combined with a ResNest200e encoder. Such an architecture contains a large number of trainable parameters that it is impractical to train it from scratch. We overcome this challenge by utilizing the Self-Distillation with No Labels (DINO) self-supervised pre-training method. Starting from the ResNest200e encoder pre-trained on ImageNet, DINO enables the model to learn rich and representative features without the need for mask annotations. Following the pre-training phase, the encoder model is integrated with the UNet++ and fine-tuned for the downstream task of segmenting the filopodia. See Sect. “Experiment details and training protocol” for implementation details.
Fig. 6.
Schematic overview of our deep learning approach for filopodia segmentation.
FiloAnalyzer’s image processing method
FiloAnalyzer incorporates an image processing pipeline similar to FiloQuant [24] and FiloDetect [22] for filopodia segmentation. Specifically, our method performs five main steps:
Adjust the gamma of the image to modify the image brightness.
Segment the cell body and filopodia with a thresholding method.
Apply erosion and dilation to get only the cell bodies.
Obtain the filopodia pixels by subtracting the cell body image from the thresholded image.
Apply a threshold on size and aspect ratio
The last step is performed to exclude noisy predictions, where the estimated filopodia are too small or do not reflect the typical aspect ratio. Our image processing method allows for dealing with images containing multiple cells. The described steps have different parameters, including the gamma value, the thresholding procedure (i.e., Hard, Otsu, or Triangular thresholding), erosion and dilation filter size and kernel, and threshold on size and aspect ratio. In our experiments, we tune such parameters to maximize accuracy and generalization, performing a cross-validation procedure. The corresponding values are inserted as the default choices in the toolbox and can be easily modified by the user exploiting the GUI.
Filopodia counting approach
Quantifying the number of filopodia in a cell is a fundamental feature of a filopodia analysis toolbox. This is, however, a task that is rendered challenging by various features of the images, especially in the case of crowded images, where crossings and overlapping happen between filopodia. As a solution, our counting approach is based on a heuristic procedure. Starting from the predicted binary mask, we perform a thinning morphological operation determining the candidate filopodia blobs via contour detection. The approximate number of filopodia is computed by counting the endpoints (ideally corresponding to the tips of filopodia) decreased by one. Our approach is similar to the one used in [22] and [24], providing a precise estimation for single filopodia and a good approximation in more complicated cases with several bifurcations. Our final output (see the bottom-right panel of Fig. 1) visually highlights such cases, where a single blob corresponds to multiple filopodia (that is, number of estimated endpoints 
for that blob). The corresponding merged filopodia are represented in yellow in the final output, while the remaining ones are visualized in purple.
Graphical user interface
In automated cell analysis, user-friendly computer tools are crucial for biologists to process data and accelerate research efficiently.
Introducing a new toolbox for filopodia segmentation and visualization should address existing limitations while providing precise results without manual parameter tuning. Its standalone nature eliminates the need for complex software environments. Developed with tkinter, an easily extendable Python library, the toolbox offers a GUI with panels to visualize the segmentation processing and results and clear statistical visualizations.
At the time of writing, the toolbox supports the following features:
Alternatives for filopodia segmentation: Supports both "FiloAnalyzer" image processing routine and Deep Learning solution. The DL solution offers faster analysis, while FiloAnalyzer allows parameter tuning per single image.
Swappable DL models: Any PyTorch model in .pth format can be loaded and used. With this choice, modularity allows the loading of potential future models with better performance
ROI selection: interactive GUI window for extracting the Region of Interest (ROI). It enables easy isolation of important elements within an image.
Batch mode: Supports processing a folder of cell images in batch. Automates a manual process, saving researchers time.
Metrics visualization: The left pane of the GUI displays metrics from segmented filopodia. Metrics include number, size, and aspect ratio presented as averages or histograms, which are the most important metrics for this task
Additional file
Acknowledgements
The authors would like to thank Dr. Annette Chan, the director of the Cell and Molecular Imaging Center at SF State University for her expertise. We would also like to thank Rocio Cisneros, Angela Lane, Madu Nzerem, Walter Orellanos, and Frederick Santana for their assistance in making the masks. S.B., V.P.P., L.B., and L.G. were funded by an NSF STC Award DBI-1548297 (Center for Cellular Construction). V.P.P. was also supported by FSE REACT-EU-PON 2014–2020, DM 1062/2021, COD. 11-G-14987-1.
Abbreviations
- ER
Endoplasmic reticulum
- WLS
Wntless
- SVM
Support vector machine
- GUI
Graphical user interface
- FCN
Fully convolutional network
- SSL
Self-supervised learning
- DDPM
Denoising diffusion probabilistic model
- EMA
Exponential moving average
- DSC
Dice score
- GT
Ground truth
- TP
True positive
- FP
False positive
- TN
True negative
- FN
False negative
- Pred
Prediction output
- ROI
Region of interest
Author contributions
V.P.P. designed the deep learning pipeline, the toolbox and the experiments. V.P.P. and R.R. implemented the toolbox. V.P.P., R.R., L.T., and R.D.V. performed the experiments and implemented the deep learning methods. L.B. and L.G. acquired, curated, and annotated the datasets. S.B. conceived and supervised the project. F.O. supervised the project. All the authors wrote and reviewed the manuscript.
Funding
S.B., V.P.P., L.B., and L.G. were funded by an NSF STC Award DBI-1548297 (Center for Cellular Construction).
Data availability
The dataset used in this work, along with the source code and executable GUI of FiloAnalyzer, is available at https://github.com/Malga-Vision/FiloAnalyzerToolbox.
Declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12859-026-06437-9.
References
- 1.Mattila PK, Lappalainen P. Filopodia: molecular architecture and cellular functions. Nat Rev Mol Cell Biol. 2008;9(6):446–54. [DOI] [PubMed] [Google Scholar]
- 2.Kyykallio H, Oikari S, Bueno Álvez M, Gallardo Dodd CJ, Capra J, Rilla K. The density and length of filopodia associate with the activity of hyaluronan synthesis in tumor cells. Cancers. 2020;12(7):1908. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Friedl P, Bröcker E-B. The biology of cell locomotion within three-dimensional extracellular matrix. Cell Mol Life Sci CMLS. 2000;57:41–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Vignjevic D, Schoumacher M, Gavert N, Janssen K-P, Jih G, Laé M, et al. Fascin, a novel target of Catenin-TCF signaling, is expressed at the invasive front of human colon cancer. Can Res. 2007;67(14):6844–53. 10.1158/0008-5472.can-07-0929. [DOI] [PubMed] [Google Scholar]
- 5.Galli LM, Santana F, Apollon C, Szabo LA, Ngo K, Burrus LW. Direct visualization of the Wntless-induced redistribution of wnt1 in developing chick embryos. Dev Biol. 2018;439(2):53–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Logan CY, Nusse R. The Wnt signaling pathway in development and disease. Annu Rev Cell Dev Biol. 2004;20(1):781–810. [DOI] [PubMed] [Google Scholar]
- 7.Takada R, Satomi Y, Kurata T, Ueno N, Norioka S, Kondoh H, et al. Monounsaturated fatty acid modification of Wnt protein: its role in Wnt secretion. Dev Cell. 2006;11(6):791–801. [DOI] [PubMed] [Google Scholar]
- 8.Wodarz A, Nusse R. Mechanisms of Wnt signaling in development. Annu Rev Cell Dev Biol. 1998;14(1):59–88. [DOI] [PubMed] [Google Scholar]
- 9.Burrus LW, McMahon AP. Biochemical analysis of murine Wnt proteins reveals both shared and distinct properties. Exp Cell Res. 1995;220(2):363–73. [DOI] [PubMed] [Google Scholar]
- 10.Bänziger C, Soldini D, Schütt C, Zipperlen P, Hausmann G, Basler K. Wntless, a conserved membrane protein dedicated to the secretion of Wnt proteins from signaling cells. Cell. 2006;125(3):509–22. [DOI] [PubMed] [Google Scholar]
- 11.Bartscherer K, Pelte N, Ingelfinger D, Boutros M. Secretion of Wnt ligands requires Evi, a conserved transmembrane protein. Cell. 2006;125(3):523–33. [DOI] [PubMed] [Google Scholar]
- 12.Goodman RM, Thombre S, Firtina Z, Gray D, Betts D, Roebuck J, Spana EP, Selva EM. Sprinter: a novel transmembrane protein required for Wg secretion and signaling 2006. [DOI] [PubMed]
- 13.Galli LM, Barnes T, Burrus LW. The use of chick embryos to study wnt activity gradients. Wnt Signaling: Methods and Protocols, 2016;69–80. [DOI] [PubMed]
- 14.Kornberg TB. Distributing signaling proteins in space and time: the province of cytonemes. Current Opinion in Genetics Dev. 2017;45:22–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Stanganello E, Hagemann AI, Mattes B, Sinner C, Meyen D, Weber S, et al. Filopodia-based wnt transport during vertebrate tissue patterning. Nat Commun. 2015;6(1):5846. [DOI] [PubMed] [Google Scholar]
- 16.Stanganello E, Scholpp S. Role of cytonemes in wnt transport. J Cell Sci. 2016;129(4):665–72. [DOI] [PubMed] [Google Scholar]
- 17.Dobos O, Horvath P, Nagy F, Danka T, Viczian A. A deep learning-based approach for high-throughput hypocotyl phenotyping. Plant Physiol. 2019;181(4):1415–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Saltukoglu D, Özdemir B, Holtmannspötter M, Reski R, Piehler J, Kurre R, et al. Plasma membrane topography governs the 3d dynamic localization of igm b cell antigen receptor clusters. EMBO J. 2023;42(4):112030. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Isensee F, Jaeger PF, Kohl SA, Petersen J, Maier-Hein KH. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. [DOI] [PubMed] [Google Scholar]
- 20.Driscoll MK, Welf ES, Jamieson AR, Dean KM, Isogai T, Fiolka R, et al. Robust and automated detection of subcellular morphological motifs in 3d microscopy images. Nat Methods. 2019;16(10):1037–44. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Quirós-Corella F, Pereira R, Corrales-Ureña YR. Microvilli semantic segmentation in microscopy images using a visual learning pipeline. In: 2023 International Conference on Machine Learning and Applications (ICMLA), 2023;1–8. IEEE
- 22.Nilufar S, Morrow AA, Lee JM, Perkins TJ. Filodetect: automatic detection of filopodia from fluorescence microscopy images. BMC Syst Biol. 2013;7:1–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Urbančič V, Butler R, Richier B, Peter M, Mason J, Livesey FJ, et al. Filopodyan: an open-source pipeline for the analysis of filopodia. J Cell Biol. 2017;216(10):3405–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Jacquemet G, Hamidi H, Ivaska J. Filopodia quantification using filoquant. Comput Optim Microscopy Methods Protocols, 2019;359–373. [DOI] [PubMed]
- 25.Eddington C, Schwartz JK, Titus MA. filovision-using deep learning and tip markers to automate filopodia analysis. J Cell Sci. 2024;137(4):jcs261274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Inc., T.M.: MATLAB Version: 9.13.0 (R2022b). https://www.mathworks.com
- 27.Schneider CA, Rasband WS, Eliceiri KW. Nih image to imagej: 25 years of image analysis. Nat Methods. 2012;9(7):671–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Chamier Lv, Laine RF, Jukkala J, Spahn C, Krentzel D, Nehme E, Lerche M, Hernández-Pérez S, Mattila PK, Karinou E, et al. Zerocostdl4mic: an open platform to use deep-learning in microscopy. BioRxiv, 2020;2020–03.
- 29.Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-assisted intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, 2015;234–241 . Springer
- 30.Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. Imagenet large scale visual recognition challenge. Int J Comput Vision. 2015;115:211–52. [Google Scholar]
- 31.Chen T, Kornblith S, Norouzi M, Hinton G. A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning, 2020;1597–1607 . PMLR
- 32.Chen T, Kornblith S, Swersky K, Norouzi M, Hinton GE. Big self-supervised models are strong semi-supervised learners. Adv Neural Inf Process Syst. 2020;33:22243–55. [Google Scholar]
- 33.He K, Fan H, Wu Y, Xie S, Girshick R. Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020;9729–9738.
- 34.Caron M, Touvron H, Misra I, Jégou H, Mairal J, Bojanowski P, Joulin A. Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021;9650–9660.
- 35.Di Via R, Odone F, Pastore VP. Self-supervised pre-training with diffusion model for few-shot landmark detection in x-ray images. In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV), 2025;3886–3896
- 36.Iakubovskii P. Segmentation Models Pytorch. GitHub 2019.
- 37.Giakoumoglou N, Giakoumoglou P. PySSL: A PyTorch implementation of Self-Supervised Learning (SSL) methods. GitHub 2023.
- 38.Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems 2020.
- 39.Manning CD, Raghavan P, Schütze H. Introduction to Information Retrieval. USA: Cambridge University Press; 2008. [Google Scholar]
- 40.Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, 2018;3–11 . Springer [DOI] [PMC free article] [PubMed]
- 41.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016;770–778.
- 42.Zhang H, Wu C, Zhang Z, Zhu Y, Lin H, Zhang Z, Sun Y, He T, Mueller J, Manmatha R, et al. Resnest: Split-attention networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022;2736–2746.
- 43.Maracani A, Pastore VP, Natale L, Rosasco L, Odone F. In-domain versus out-of-domain transfer learning in plankton image classification. Sci Rep. 2023;13(1):10443. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Walt S, Schönberger JL, Nunez-Iglesias J, Boulogne F, Warner JD, Yager N, et al. contributors: scikit-image: image processing in Python. PeerJ. 2014;2:453. 10.7717/peerj.453. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015;3431–3440. [DOI] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The dataset used in this work, along with the source code and executable GUI of FiloAnalyzer, is available at https://github.com/Malga-Vision/FiloAnalyzerToolbox.