Model-Free Framework for Pseudo-Label Generation

A Resource-Efficient and Transparent Model-Free Framework for Pseudo-Label Generation in Material CT Scans

This repository hosts the source code for the model-free pseudo-label generation from CT scanned images, as discussed in the paper "A Resource-Efficient and Transparent Model-Free Framework for Pseudo-Label Generation in Material CT Scans." The proposed framework establish a transparent approach as a highly effective tool for annotation bootstrapping and microstructural quantification in resource-limited environments.

The Framework

The proposed framework has three key sections as displayed in the figure below.

Figure: The three key sections of the proposed framework for pseudo-label generation.

In step one we use traditional image processing techniques to generate labels and outlines from grayscale images (e.g. CT scans).

In step two we dilate the outlines and use a custom threshold-based BFS algorithm for region growing. This custom algorithm creates masks for regions of interest (ROIs) in the images.

In step three we use the generated masks to create pseudo-labels for the original grayscale images. These pseudo-labels can then be used for training machine learning models for downstream segmentation tasks.

Experiments

Experiments on different datasets prove the efficiency of the proposed framework. The proposed approached produced comparable results to state-of-the-art foundation models running on GPU, while using only CPU resources and taking only half the time.

Sand Grain Dataset

Experiment on the APS Sand Grain Dataset (Detail ↗) shows that the proposed framework can generate high-quality pseudo-labels for sand grain images.

Figure: The original grayscale CT scan image (left), the generated pseudo-label using our method (middle), and the generated pseudo-label using zero-shot SAM (right).

The pseudo-labels were generated 2x faster on CPU compared to the zero-shot SAM model running on GPU, while producing comparable results. Downstream segmentation tasks from the generated pseudo-labels also produced comparable results to the zero-shot SAM model. We trained YOLOv11 model using both the pseudo-labels to compare the results.

Table: Comparison of the proposed model-free framework with zero-shot SAM on the APS Sand Grain Dataset.

Method	Parameters(Mi)	CPU	GPU	RAM	Time(sec)
Model-free	$-$	$\checkmark$	$\times$	$\checkmark$	$15/im$
SAM	$80.9$	$\checkmark$	$\checkmark$	$\checkmark$	$28/im$

Dendrites Dataset

Experiment on the Dendrites Dataset (Detail ↗) shows that the proposed framework can generate high-quality pseudo-labels on even lower contrast images.

Figure: The original grayscale CT scan image (left), the generated pseudo-label using our method (middle), and the generated pseudo-label using zero-shot SAM (right).

Table: Comparison of the proposed model-free framework with zero-shot SAM on the Dendrites Dataset.

Method	Accuracy	IoU	F1
SAM	$98.63%$	$10.74%$	$19.39%$
Model-free	$99.12%$	$63.19%$	$77.44%$

Microstructure Quantification

The proposed framework performed well in microstructure quantification tasks, producing comparable results to the zero-shot SAM model. This experiment proves the effectiveness of the framework in resource-limited materials science experiments.

Figure: Effective circular diameter (ECD) comparison between the proposed model-free framework and zero-shot SAM on sand grains.

Dermatology Dataset

Beyond materials dataset and grayscale CT scans, the algorithm can be tuned to work on a diverse environment. A quick experiment on Diverse Dermatology Images (DDI) dataset (Detail ↗), produced appropriate masks for ROIs on diverse skin tones.

Figure: Pseudo-label generation on sample DDI data.

Directory Structure

The directory structure of the repository is as follows:

├── figures/                        # Contains the figures used in the paper and README
|── generate_pseudo_label.ipnyb     # A Jupyter notebook for generating pseudo-labels from CT scans
|── PseudoLabelAnnotation.py        # A Python class implementation of the proposed framework for pseudo-label generation
|── utility.py                      # A utility file containing helper functions used in the class
|── README.md                       # This README file

Citation

The complete work is currently under review for publication. If you find this work useful in your research, please consider citing:

@article{Rahman2026cv,
  title={A Resource-Efficient and Transparent Model-Free Framework for Pseudo-Label Generation in Material CT Scans},
  author={Rahman, Ashiqur; Seethi, Venkata Devesh Reddy; Yunker, Austin; Kettimuthu, Rajkumar; and Alhoori, Hamed;},
  year={2026}
}