PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ashiqur Rahman, Venkata Devesh Reddy Seethi, Austin Yunker, Rajkumar Kettimuthu, Hamed Alhoori

arXiv:2610.28974v1cs.CV

TL;DR

Scarce expert labels and costly, opaque deep-learning pseudo-labeling complicate material CT segmentation, while isolated classical techniques can be fragile. TEMPL combines model-free image-processing stages to produce polygonal annotations directly from grayscale scans. It achieves 63.19% IoU versus 10.74% for zero-shot SAM on the Dendrites dataset, and its limitations include sensitivity to image artifacts and complex interconnected morphologies.

  • Problem

    Material CT segmentation is constrained by scarce expert annotations, while deep-learning pseudo-labeling can be computationally expensive and difficult to interpret.

    Source: §1 · p. 2

  • Method

    TEMPL combines classical image-processing stages, including thresholding, dilation, and custom region-growing, to turn grayscale CT scans into polygonal pseudo-labels.

    Source: §1 · p. 2, §4 · p. 6, Step 1 - Dilation · p. 8

  • Results

    63.19% IoU without labeled data versus 10.74% for zero-shot SAM on the Dendrites dataset.

    Source: §6 · p. 20

  • Takeaways & Limitations

    Models trained on TEMPL pseudo-labels achieved detection and instance-segmentation performance comparable to models trained on SAM annotations, with pseudo-label generation nearly twice as fast on a standard CPU.

    Source: §8 · p. 22

    TEMPL is sensitive to CT artifacts such as beam hardening and electronic noise, and may struggle with interwoven fibers or interconnected pores without specialized region-growing modifications.

    Source: §7 · p. 21

Abstract

from arXiv · show

The scarcity of large annotated datasets limits deep learning (DL) segmentation in material CT analysis. DL-based pseudo-labeling can help, but it is often computationally expensive and hard to interpret. We introduce a multi-stage, model-free framework that generates pseudo-labels directly from grayscale CT scans. It combines thresholding, morphological dilation, and a custom breadth-first search region-growing algorithm to produce polygonal annotations without pre-trained weights, labeled data, or a GPU. We compare it against the zero-shot Segment Anything Model (SAM), a fully supervised 3D U-Net, and classical baselines. On a dendrite dataset with ground truth, our method reaches 63.19% IoU versus 10.74% for zero-shot SAM, with four times the instance recall at 0.5 IoU (33.3% versus 8.3%). The supervised 3D U-Net serves as a reference ceiling that requires expert labels. On a silica sand specimen, YOLO models trained on our labels performed comparably to models trained on SAM labels in a qualitative comparison. Across 1,053 extracted grains, the median equivalent diameter was 163.8 micrometers, near the specimen's known 106 to 150 micrometers sieve range, compared with 219.5 micrometers for SAM, whose labels merge adjacent particles and add spurious detections. A sensitivity analysis shows a stable operating range for the key threshold and a sharp, predictable failure point linked to phase contrast. Our method needs 15s per image on a standard CPU versus 28s for SAM on a GPU. These results support our transparent approach as a tool for annotation bootstrapping and microstructural quantification in resource-limited settings.

1 Introduction

Material CT segmentation is constrained by scarce expert labels, while deep-learning pseudo-labeling can be costly and difficult to interpret and isolated classical methods can be fragile. TEMPL combines model-free image-processing stages to produce polygonal labels, with evaluation indicating stronger annotation signals than zero-shot SAM and utility for downstream training.

1 Introduction

  • Expert ground truth for the dendritic dataset takes about 20 minutes per image slice, a cost that scales across volumetric CT scans.Annotation expense and data variability complicate model deployment and generalization.Source: §1 · p. 2
  • Standalone thresholding, edge detection, and clustering are resource-efficient but fragile under image variation, while deep-learning pseudo-labeling is costly and hard to explain.The paper identifies no widely adopted combined framework tailored to grayscale material CT intensity profiles and morphologies.Source: §1 · p. 2
  • TEMPL combines adaptive thresholding, morphological operations, and custom region-growing to generate polygonal annotations from material CT scans.Its region-growing uses automated local seed identification and BFS-based propagation for discontinuous boundaries and heterogeneous intensities.Source: §1 · p. 2
  • Models trained on TEMPL labels achieved detection and instance-segmentation performance comparable to models trained on SAM labels, while TEMPL ran nearly twice as fast on CPU without GPU resources.The evaluation also reports that zero-shot SAM hallucinates background instances and merges adjacent particles.Source: §1 · p. 3

2 Related Works

Prior segmentation and pseudo-labeling work spans supervised deep learning, unsupervised models, and classical methods, but the paper identifies unresolved efficiency, generalization, and robustness challenges for grayscale material CT. TEMPL is positioned as an efficient, interpretable combination of model-free techniques for this setting.

2.1 Image Segmentation and Object Detection

  • Deep-learning segmentation includes supervised architectures such as U-Net and instance-detection systems such as Mask R-CNN, YOLO, and zero-shot SAM.The literature also describes their use in medical imaging and industrial defect detection.Source: §2.1 · p. 3
  • Expert labeling remains laborious, expensive, and dependent on domain expertise in specialized areas including materials science and medical imaging.This data bottleneck motivates unsupervised and semi-supervised approaches.Source: §2.1 · pp. 3–4

2.2 Unsupervised Models

  • Unsupervised and semi-supervised learning use unlabeled data, but accuracy and generalizability without prohibitive computational cost remain central challenges.Pseudo-label generation complexity and cross-domain generalizability remain open questions.Source: §2.2 · p. 4
  • Unsupervised DL methods include W-Net, CUTS, U2Seg, and UnSAM, which use approaches such as autoencoding, clustering, or decomposition to generate pseudo-labels.CUTS is described as performing comparably to foundation models without large labeled datasets.Source: §2.2 · p. 4

2.3 Pseudo-label Generation

  • Existing seeded region-growing formulations rely on manually selected seeds and conventional connectivity assumptions, while materials-CT applications may require prior phase specification.These constraints limit use on unlabeled, heterogeneous monochromatic CT with discontinuous phase boundaries and within-particle intensity variation.Source: §2.3 · p. 4
  • No single classical technique reliably handles discontinuous boundaries, heterogeneous intensity profiles, and noise in material CT, motivating a composite pipeline.Thresholding, Watershed, superpixels, and active contours each have distinct noise, low-contrast, initialization, or computational limitations.Source: §2.3 · p. 5
  • Self-training and confidence or consensus strategies aim to improve pseudo-label fidelity, while single-model approaches can reinforce prediction errors through confirmation bias.Teacher-student and mutual-training frameworks are described as countermeasures.Source: §2.3 · p. 5
  • TEMPL combines heuristic approaches into a rapid, transparent model-free benchmark for image segmentation.The paper positions it as a cost-effective alternative to computationally intensive foundation-model pseudo-label generation.Source: §2.3 · p. 5

3 Dataset

The study uses unlabeled silica-sand CT images for pseudo-label generation and a separate dendrite dataset with ground truth for quantitative validation. The primary dataset is a resized grayscale image collection, while the validation data are divided into two-dimensional slices.

3 Dataset

  • The study separates unlabeled sand-grain images for pseudo-label generation from a dendrite dataset with ground-truth labels for quantitative validation.The validation volume was partitioned into 2D z-slices for solid- and liquid-phase segmentation.Source: §3 · p. 5, §3.2 · p. 6

3.1 Primary Dataset

  • The final silica-sand dataset contains 1,192 grayscale images at 769 × 770 resolution, capturing grains with diverse intensity profiles.Slices without discernible grains were excluded and the remaining images were resized for computational efficiency.Source: §3.1 · p. 5

3.2 Validation Dataset

  • The Dendrites validation set was acquired using synchrotron-based 4D X-ray CT at Argonne National Laboratory and provides ground truth for evaluating model-free outputs.Fig. 1b shows a representative validation sample.Source: §3.2 · p. 6

4 Method

TEMPL generates polygonal pseudo-labels from grayscale CT scans through a staged pipeline that combines classical image processing with dilation, custom region growing, and mask refinement. Each stage addresses a limitation of the preceding step and exposes tunable parameters so errors can be traced to their source.

4 Method

  • TEMPL converts preliminary grayscale-image outlines into polygonal annotations through image processing, dilation, region growing, and mask refinement.The pipeline has two phases: generating preliminary outlines, then enhancing them into final annotations.Source: §4 · p. 6

4.1 Image Processing

  • Otsu thresholding and connected-component labeling create initial grain regions, but global thresholding can fragment grains at low-contrast boundaries.Regions smaller than 10×10 pixels are discarded; dilation and region growing address the fragmentation.Source: §4.1 · pp. 6–8

4.2 Label Generation

  • The four-step label-generation sequence closes boundary gaps, merges contiguous regions, fills object interiors, then cleans masks and converts them to polygons.The stages target limitations of preceding steps and expose tunable parameters for tracing errors.Source: §4.2 · p. 8

Step 1 - Dilation

  • A 3×3 dilation expands foreground outlines by one pixel in each direction, closing minor gaps while limiting merges between neighboring grains.Larger kernels bridge wider gaps but raise the risk of merging adjacent particles; smaller kernels may leave gaps unresolved.Source: Step 1 - Dilation · p. 8

Step 2 - Create Labels

  • Full-connectivity component labeling merges regions that touch at an edge or corner after dilation, including fragments meeting at one corner pixel.The resulting label map and regions are compared with those from the initial processing stage.Source: Step 2 - Create Labels · p. 8

Step 3 - Region Growing

  • The region-growing stage identifies an interior seed and uses breadth-first search to fill the object mask, with a 2-pixel neighborhood reach to limit leakage.A pixel is included when it is in bounds, unvisited, and within intensity threshold τ of the seed value; the search stops when no eligible neighbors remain.Source: Step 3 - Region Growing · pp. 9–10
  • The custom region-growing algorithm uses raster-scan seed identification, symmetric BFS expansion, and a restricted 5×5 neighborhood to address heterogeneous intensities and boundary leakage.The local seed requires no prior knowledge of phase count or intensity ranges; the 2-pixel reach encodes an assumption about gaps between adjacent particles.Source: Step 3 - Region Growing · p. 10

Step 4 - Refinement

  • Refinement removes small foreground fragments and holes below 150 pixels, relabels connected regions as instances, then traces contours into polygon annotations.The 150-pixel threshold matches the one used in the matched-scale Watershed comparison.Source: Step 4 - Refinement · pp. 10–11

5 Experiment

The experiments assess TEMPL labels qualitatively through downstream segmentation and quantitatively against established ground-truth annotations. They use a CUDa-enabled workstation with an Intel Core i9 processor, one GPU, and 64GB of RAM.

5 Experiment

  • The evaluation combines qualitative and quantitative assessments of labels generated by the framework.The quantitative assessment compares pseudo-labels with established ground-truth annotations, while the qualitative assessment examines downstream segmentation utility.Source: §5 · p. 11
  • Two YOLOv11 models were trained separately using TEMPL labels and labels from an established deep-learning approach for a downstream segmentation comparison.The comparison evaluates the practical utility of TEMPL labels against a more resource-intensive alternative.Source: §5 · p. 11
  • Figure 6 presents label-generation attempts from W-Net and CUTS.Source: Figure 6 · p. 12

5.1 Experimental Setup

  • Experiments ran on a CUDa-enabled workstation with an Intel Core i9 3.2GHz processor, one 12GB GPU, and 64GB RAM, using Python 3.11.Source: §5.1 · p. 12

5.2 Qualitative Evaluation

The evaluation compares TEMPL with SAM-generated labels, then tests models trained on each label set for object detection and instance segmentation. The methods are qualitatively comparable, with different error patterns and resource costs.

5.2 Qualitative Evaluation

  • W-Net and CUTS generated noisy or indistinct sand-grain masks and were unsuitable for the grayscale CT images.CUTS marginally improved grain-interior separation, but its boundaries and masks remained insufficient.Source: §5.2 · p. 12

Comparison of Pseudo-Labels

  • TEMPL and zero-shot SAM produce qualitatively comparable pseudo-labels, but their noise differs in type and explainability.TEMPL's motion-artifact noise can be mitigated by parameter tuning; SAM's overlapping boxes and segmentation artifacts are harder to explain or correct.Source: Comparison of Pseudo-Labels · p. 13, Figure 7 · p. 13, Figure 8 · p. 14
  • 15 seconds per image on a CPU for TEMPL compares with 28 seconds per image for SAM using a 12GB GPU.Table 1 compares label-generation resources and shows faster throughput for the model-free approach without GPU resources.Source: Comparison of Pseudo-Labels · p. 13, Table 1 · p. 14

Object Detection

  • Models trained on TEMPL and SAM labels perform comparably in qualitative object detection, with over-segmentation and redundant overlapping boxes, respectively.The reported error patterns correspond to noise in each training-label set.Source: Object Detection · p. 14, Figure 9 · p. 15

Instance Segmentation

  • SAM-trained models produce more complete instance segmentations, while TEMPL-trained models competitively capture most grain shapes but leave some internal voids.Those voids correspond to small gaps in TEMPL's pseudo-labels.Source: Instance Segmentation · p. 15
  • TEMPL errors are traceable and can be mitigated by tuning parameters, whereas correcting SAM's less predictable failure modes is non-trivial.The comparison frames accuracy, computational efficiency, explainability, and iterative refinement as trade-offs.Source: Instance Segmentation · p. 15

5.3 Quantitative Evaluation

On the Dendrites dataset, TEMPL outperforms zero-shot SAM on pixel-level overlap and instance recall, while the supervised 3D U-Net is presented as a reference ceiling requiring expert labels. Watershed performance is similar in some pixel-level values but requires fine-tuning and fragments regions.

5.3 Quantitative Evaluation

  • Watershed required excessive fine-tuning and fragmented true regions into smaller pieces described as unusable for downstream training.Its over-segmentation is visible in Fig. 11b.Source: §5.3 · pp. 16–17, Figure 11 · p. 16
  • 63.19% IoU and 77.44% F1 for TEMPL compare with 10.74% IoU and 19.39% F1 for zero-shot SAM.Pixel accuracy above 98% for both is described as misleading because background occupies much of the CT scans.Source: §5.3 · p. 16, Table 2 · p. 17
  • The supervised 3D U-Net reached a maximum 95.56% IoU, serving as a reference ceiling because it requires expert-annotated training data.TEMPL achieved 63.19% IoU without training or manual labeling overhead; the passage reports about 20 minutes of expert labor per image slice for ground truth.Source: §5.3 · p. 17
  • 33.33% recall at the 50% threshold for TEMPL is four times SAM's 8.33% recall.SAM also had 100% precision at both 50% and 75% overlap thresholds, but low recall on the complex dendritic structures.Source: §5.3 · p. 17, Table 3 · p. 18

5.4 Microstructural Quantification

Across 100 randomly selected sand slices, TEMPL yielded 1,053 filtered grains with a median ECD near the specimen's sieve range. Its ECD distribution differed significantly from SAM's, whose labels were reported to merge adjacent particles and add spurious detections.

5.4 Microstructural Quantification

  • 1,053 valid grains extracted by TEMPL had a median ECD of 163.8µm, compared with 219.5µm for SAM.The specimen's sieve range was 106–150µm; the TEMPL median was described as close to its 150µm upper limit within expected 2D stereological sampling variation.Source: §5.4 · p. 18, Figure 12 · p. 18
  • The ECD distributions differed significantly (KS = 0.215, p < 0.0001), with the KS statistic marking the maximum separation between their cumulative distributions.The reported interpretation is that SAM merges adjacent particles into artificially large structures and adds detections from background noise, while TEMPL provides physically grounded quantification.Source: §5.4 · pp. 18–19, Figure 13 · p. 19

5.5 Parameter Sensitivity Analysis

The intensity-similarity threshold has a stability plateau below baseline and a sharp deterioration above it. The plateau width and transition point depend on the Dendrites dataset's phase contrast and may differ for other image characteristics.

5.5 Parameter Sensitivity Analysis

  • From τ −4 through τ −1, all segmentation metrics remain unchanged relative to baseline.The reported stability spans four threshold units and indicates that precise calibration is unnecessary within this regime for similar material datasets.Source: §5.5 · pp. 19–20
  • At τ +1 and τ +2, precision drops below 7% while pixel-level recall remains near unity; at τ +3 and τ +4, no valid object masks remain after refinement.The passage attributes this deterioration to threshold leakage beyond the solid-liquid interface into the background.Source: §5.5 · p. 20
  • The plateau width and transition point are specific to Dendrites; lower phase contrast or more heterogeneous internal intensities may narrow the stable range and require more careful threshold selection.The threshold's reported transition is linked to the image's solid-liquid intensity contrast.Source: §5.5 · p. 20

6 Discussion

TEMPL offers a fast, transparent alternative to DL-based pseudo-labeling, with stronger overlap than zero-shot SAM and practical value for downstream analysis and prototyping. Its labels also support downstream YOLO training comparable to SAM labels in the reported qualitative assessment.

6 Discussion

  • 63.19% IoU without labeled data, versus 10.74% for zero-shot SAM; the supervised 3D U-Net’s 95.56% is a reference ceiling requiring expert labels.The paper frames TEMPL as a viable starting point for hybrid pipelines.Source: §6 · p. 20
  • YOLO trained on TEMPL labels performed remarkably comparably to YOLO trained on zero-shot SAM labels in downstream segmentation.On a 100-slice batch, TEMPL also extracted a filtered grain population that closely tracked the specimen’s physical sieve limits.Source: §6 · p. 21
  • TEMPL’s transparent predictions can be traced to tunable parameters, supporting interpretable error diagnosis and domain-specific adaptation.Source: §6 · p. 21
  • 15 seconds per image on a CPU versus 28 seconds for SAM with GPU acceleration makes TEMPL faster while requiring less capable hardware.Source: §6 · p. 21
  • TEMPL can establish rapid-prototyping baselines and provide labels for pretraining before fine-tuning on a small expert-annotated dataset.Source: §6 · p. 21

7 Limitations and Future Works

TEMPL’s performance is constrained by image quality and object morphology, and its parameter stability varies with phase contrast. The downstream YOLO comparison is also limited by a two-image test split.

7 Limitations and Future Works

  • CT artifacts such as beam hardening and electronic noise challenge intensity-based segmentation, while interwoven fibers and interconnected pores may require modified region-growing logic.Source: §7 · p. 21
  • The threshold’s stability plateau may narrow with lower phase contrast, heterogeneous intensities, or beam-hardening artifacts, requiring careful selection rather than one fixed threshold across datasets.The transition remains tied to image phase contrast and can be identified through histogram inspection; dilation kernels may also need adjustment for finer features or thinner boundaries.Source: §7 · pp. 21–22
  • The Dendrites downstream detection test split contains only two images, too few for a statistically meaningful comparison of YOLO models trained on TEMPL versus SAM labels.Source: §7 · p. 22

8 Conclusion

The paper concludes that its model-free framework produces pseudo-labels with stronger overlap, recall, and instance integrity than zero-shot foundation-model and Watershed alternatives. It also reports comparable downstream model performance and practical value for rapid prototyping and microstructural quantification.

8 Conclusion

  • TEMPL achieves higher overlap and recall than the zero-shot foundation model, with stronger instance-level structural integrity than the foundation model or Watershed at matched noise-removal thresholds.Source: §8 · p. 22
  • Models trained with TEMPL pseudo-labels achieved object detection and instance segmentation performance comparable to models trained with SAM annotations, while labels took nearly half the per-image time on a standard CPU.SAM-based methods required twice the duration despite GPU acceleration.Source: §8 · p. 22
  • TEMPL supports rapid prototyping, performance baselines, domain-specific adaptation, error analysis, and grain count, morphology, and size-distribution quantification.Source: §8 · p. 22

Related papers

Loading 2610.28974v1…