Controlled Benchmarking and Robustness Analysis of Classical Machine-Learning Approaches for Synthetic CT Sample Segmentation

Reliable delineation of a specimen from its surrounding background is a prerequisite for quantitative analysis of computed tomography (CT) images. This work is positioned as a controlled benchmarking and robustness study of classical machine-learning approaches for binary sample–background segmentation in synthetic CT-like images, rather than as a demonstration of validated industrial CT applicability. Five supervised classifiers—Logistic Regression (LR), Gaussian Naive Bayes (GNB), linear Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF)—are compared using a synthetic dataset of 300 grayscale images (256 × 256 pixels) with corresponding binary reference masks. Three pixel-level feature representations are examined: intensity alone; intensity with normalized spatial coordinates; and a five-dimensional representation comprising intensity, two spatial coordinates, local mean, and local standard deviation calculated over a 5 × 5 neighbourhood. The dataset is divided into training, validation, and testing subsets containing 210, 45, and 45 images, respectively. Under the constructed synthetic conditions, the DT classifier using the complete feature set achieves the highest reported mean test accuracy (0.9998), Dice coefficient (0.9997), and intersection over union (IoU; 0.9995). Sensitivity experiments reveal marked performance degradation as additive Gaussian noise increases and image contrast decreases, whereas increasing the training-set size produces only marginal gains. The principal contribution is a controlled comparison of classical pixel-wise classifiers, an assessment of feature-engineering effects, and a robustness analysis under specified synthetic perturbations. The results characterize model behaviour within the defined synthetic data-generating process; they do not establish performance on experimental or industrial XCT data. Validation using independent, representative XCT scans is a necessary subsequent step.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11740273/v1
Primary Topic
Medical Image Segmentation Techniques
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Controlled Benchmarking and Robustness Analysis of Classical Machine-Learning Approaches for Synthetic CT Sample Segmentation

Mayank Goswami, Musunuri Rajani Priya
4 links
Research Square
Medical Image Segmentation Techniques
preprint

Controlled Benchmarking and Robustness Analysis of Classical Machine-Learning Approaches for Synthetic CT Sample Segmentation

Mayank Goswami, Musunuri Rajani Priya
preprint en

Abstract

Reliable delineation of a specimen from its surrounding background is a prerequisite for quantitative analysis of computed tomography (CT) images. This work is positioned as a controlled benchmarking and robustness study of classical machine-learning approaches for binary sample–background segmentation in synthetic CT-like images, rather than as a demonstration of validated industrial CT applicability. Five supervised classifiers—Logistic Regression (LR), Gaussian Naive Bayes (GNB), linear Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF)—are compared using a synthetic dataset of 300 grayscale images (256 × 256 pixels) with corresponding binary reference masks. Three pixel-level feature representations are examined: intensity alone; intensity with normalized spatial coordinates; and a five-dimensional representation comprising intensity, two spatial coordinates, local mean, and local standard deviation calculated over a 5 × 5 neighbourhood. The dataset is divided into training, validation, and testing subsets containing 210, 45, and 45 images, respectively. Under the constructed synthetic conditions, the DT classifier using the complete feature set achieves the highest reported mean test accuracy (0.9998), Dice coefficient (0.9997), and intersection over union (IoU; 0.9995). Sensitivity experiments reveal marked performance degradation as additive Gaussian noise increases and image contrast decreases, whereas increasing the training-set size produces only marginal gains. The principal contribution is a controlled comparison of classical pixel-wise classifiers, an assessment of feature-engineering effects, and a robustness analysis under specified synthetic perturbations. The results characterize model behaviour within the defined synthetic data-generating process; they do not establish performance on experimental or industrial XCT data. Validation using independent, representative XCT scans is a necessary subsequent step.

Research Square
Indian Institute of Technology Roorkee (IN)
Indian Institute of Technology Roorkee
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon