Effects of training dataset size and model architecture on deep learning instance segmentation for bubble measurement in two-phase flow

Accurate bubble segmentation is essential for extracting quantitative information from high-speed images of bubbly two-phase flows. However, variations in bubble size, deformation, overlap, and image quality remain challenging for conventional image-processing methods. This study evaluates deep-learning-based bubble instance segmentation using standalone models and sequential detection–segmentation pipelines. Seven nested datasets ranging from 500 to 2,000 images in increments of 250 were developed from three sources, namely, experimental images, semi-organic images assembled from bubbles produced by a generative adversarial network (GAN), and fully synthetic images. Models trained on these datasets were evaluated on an independent 600-image benchmark containing 6,689 bubble instances. The standalone models included Mask R-CNN, RF-DETR-Seg Small, Large, and 2XL, and YOLO26x-Seg, while the sequential approach combined YOLO26 and RF-DETR detectors with SAM 2.1 or SAM 3. Performance was evaluated using conventional segmentation metrics together with bubble-specific diameter, centroid, and count errors, as well as inference latency. RF-DETR-Seg-2XL achieved the highest standalone AP@[0.50:0.95] of 0.887, with a recall of 0.973, F1 score of 0.983, and mean matched intersection over union (IoU) of 0.935. The highest sequential average precision (AP) was 0.821 for RF-DETR-2XL + SAM 3. Detector architecture produced larger differences in sequential-pipeline performance than the choice of SAM version, while the standalone models provided higher segmentation accuracy and substantially lower inference latency. Domain-specific evaluation further showed lower performance on experimental images than on semi-organic and synthetic images, particularly at stringent IoU thresholds. These results demonstrate the effectiveness of RF-DETR-based standalone instance segmentation for quantitative bubble-image analysis.

Authors

Institutions

Publication Details

Journal
Frontiers in Nuclear Engineering
Published
2026-11-19
DOI
https://doi.org/10.3389/fnuen.2026.2028300
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Effects of training dataset size and model architecture on deep learning instance segmentation for bubble measurement in two-phase flow

Khondokar Nushaer Abrar, Yang Liu
Frontiers in Nuclear Engineering
Advanced Neural Network Applications
article

Effects of training dataset size and model architecture on deep learning instance segmentation for bubble measurement in two-phase flow

Khondokar Nushaer Abrar, Yang Liu
article en

Abstract

Accurate bubble segmentation is essential for extracting quantitative information from high-speed images of bubbly two-phase flows. However, variations in bubble size, deformation, overlap, and image quality remain challenging for conventional image-processing methods. This study evaluates deep-learning-based bubble instance segmentation using standalone models and sequential detection–segmentation pipelines. Seven nested datasets ranging from 500 to 2,000 images in increments of 250 were developed from three sources, namely, experimental images, semi-organic images assembled from bubbles produced by a generative adversarial network (GAN), and fully synthetic images. Models trained on these datasets were evaluated on an independent 600-image benchmark containing 6,689 bubble instances. The standalone models included Mask R-CNN, RF-DETR-Seg Small, Large, and 2XL, and YOLO26x-Seg, while the sequential approach combined YOLO26 and RF-DETR detectors with SAM 2.1 or SAM 3. Performance was evaluated using conventional segmentation metrics together with bubble-specific diameter, centroid, and count errors, as well as inference latency. RF-DETR-Seg-2XL achieved the highest standalone AP@[0.50:0.95] of 0.887, with a recall of 0.973, F1 score of 0.983, and mean matched intersection over union (IoU) of 0.935. The highest sequential average precision (AP) was 0.821 for RF-DETR-2XL + SAM 3. Detector architecture produced larger differences in sequential-pipeline performance than the choice of SAM version, while the standalone models provided higher segmentation accuracy and substantially lower inference latency. Domain-specific evaluation further showed lower performance on experimental images than on semi-organic and synthetic images, particularly at stringent IoU thresholds. These results demonstrate the effectiveness of RF-DETR-based standalone instance segmentation for quantitative bubble-image analysis.

Frontiers in Nuclear EngineeringVol. 5
Virginia Tech (US)
Advanced Neural Network Applications
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