Anchored Fusion for Automated Airway Segmentation in Pediatric and Adult CT

Huseyin Umut Isik; Ege Mert Balcik; Duygu Atasoy; Asutosh Sahu; Seyda Ertekin; Sila Kurugol

eess.IVOct 28, 2026

Anchored Fusion for Automated Airway Segmentation in Pediatric and Adult CT

Authors: Huseyin Umut Isik, Ege Mert Balcik, Duygu Atasoy, Asutosh Sahu, Seyda Ertekin, Sila Kurugol

Organizations: Department of Computer Engineering, METU, Ankara, Turkey · Department of Radiology, Boston Children’s Hospital and Harvard Medical School, Boston, MA, USA · METU-DTX Digital Transformation and Innovation Center, Ankara, Turkey

Abstract

Airway segmentation from chest CT enables branch-level measurement of airway size and growth, but peripheral bronchi remain hard to segment, particularly in children, whose airways are smaller and for whom few annotated scans exist. Different models recover complementary branches of the airway tree. Their union improves overall tree coverage but also introduces disconnected non-airway objects. We propose anchored fusion (A-Fuse), which takes the voxel-wise union of TfeNet, a retrained MEDPSeg and a reimplemented AirMorph and retains the connected component with greatest overlap with the AirMorph mask. On 30 held-out adult CTs from the ATM'26 training set, A-Fuse reaches a Dice score of 93.24% and the highest branch detection rate among compared methods (92.05%). On nine pediatric scans, it reaches 94.15% Dice and 68.79% branch detection. A-Fuse retains the airway tree in all 39 test scans, whereas keeping the largest component selects a non-airway object in four. A-Fuse is deployed as an automated web application and we release all code, trained weights and an open-source ITK-SNAP extension for 3D tubular segmentation editing.

Figures & tables

Figure 1: Overview of our method. Segmentation (top to bottom): AirMorph-R produces $A$ ; MEDPSeg-R an d TfeNet contribute their raw outputs $\tilde{M}$ and $\tilde{T}$ . Their union $U=A\cup\tilde{M}\cup\tilde{T}$ is reduced to the connected component with greatest overlap with $A$ , yielding the airway mask $F$ (Eq. 1 ). Confidence (right): each voxel’s prediction pattern across AirMorph-R, MEDPSeg-R, and TfeNet is mapped to its precision on the adult validation set, providing voxel-level and branch-level confidence without modifying $F$ .

Figure 2: Component selection in the 0.9-year-old infant. Top: model outputs after their standard post-processing, (a) AirMorph-R ( $A$ ), (b) MEDPSeg-R ( $M$ ) and (c) TfeNet ( $T$ ); (b) and (c) select non-airway components. Bottom: (d) union of $A$ with the raw outputs $\tilde{M}$ and $\tilde{T}$ , (e) its largest component and (f) anchored selection $F$ (enlarged). Panels (a)–(c) and (e) share one scale. Dice against the ground truth and mask volume are given below each panel.

	Adult ATM'26 ( $n=30$ )						Pediatric CT ( $n=9$ )
Strategy	DSC	clDice	TLD	BND	Sens.	HD95	DSC	clDice	TLD	BND	Sens.	HD95
AirMorph-R ( $A$ )	92.06	90.29	94.30	90.81	94.92	2.32	87.70	87.85	76.10	59.81	78.27	2.80
MEDPSeg-R ( $M$ )	93.84	87.59	79.29	68.38	93.16	2.11	78.94	59.39	41.63	29.51	71.14	21.27
TfeNet ( $T$ )	84.44	80.86	74.94	64.96	81.64	19.06	70.06	66.45	55.90	41.87	63.80	26.54
$A\cup M$	93.17	90.11	94.55	91.27	97.84	1.80	86.64	86.66	77.63	61.01	86.53	12.37
$A\cup T$	89.83	87.37	95.03	91.86	97.43	15.46	84.32	85.19	80.37	65.97	83.62	19.12

Table 1: Segmentation results (mean; % except HD95 in mm).

A

M

T

: outputs after each model’s own post-processing;

\tilde{M}

\tilde{T}

: raw outputs. Anchored: component overlapping

A

most; Ours: anchored

A\cup\tilde{M}\cup\tilde{T}

. Bold/underline: best/second-best per cohort.

Metric	Strategy	1	2	3	4	5	6	7	8	9	10	WT
DSC	AirMorph-R	95.15	98.29	97.32	95.06	90.86	86.00	81.06	81.29	80.23	75.74	92.06
	MEDPSeg-R	98.29	98.26	97.10	95.12	91.74	87.33	80.63	79.93	77.97	73.99	93.84
	TfeNet	87.57	87.89	87.06	85.58	83.44	80.37	75.48	77.52	76.00	70.92	84.44
	A-Fuse	98.44	98.13	96.98	94.72	90.46	85.62	80.50	80.88	79.76	75.27	93.24
clDice	AirMorph-R	95.68	100.00	99.88	99.57	95.73	92.14	87.50	88.13	87.28	83.66	90.29
	MEDPSeg-R	99.86	99.96	99.61	99.20	95.23	90.80	83.82	83.51	82.71	79.51	87.59

Table 2: Adult results per generation (mean, %). Generation 1 is the trachea; generations 8–10 are present in 29 of the 30 volumes. WT: whole tree. Members use their own post-processing, as in Table 1 . Bold indicates the best values for each metric and generation.

Figure 3: A 1.7-year-old child (top) and an adult ATM test case (bottom). (a, d) A-Fuse segmentation $F$ colored by vote-pattern confidence (red low, green high). (b, e) AirMorph-R and (c, f) A-Fuse ( $F$ ) in blue, with missed ground truth branches ( $<80\%$ centerline coverage) in orange. Arrows mark branches recovered by A-Fuse. Missed branches decrease from 109 to 66 of 292 in the child and from 50 to 36 of 293 in the adult; BND increases from 62.7% to 77.4% and 82.9% to 87.7%, respectively. This illustrates improved peripheral branch recovery in pediatric and adult CT.

Image Segmentation

Image segmentation assigns pixels or voxels to objects and regions, supporting measurement, scene understanding, and decisions across image domains.

See all 406 papers

BifDet: A 3D Bifurcation Detection Dataset for Airway-Tree Modeling
Apr 27, 2026Ali Keshavarzi, Quentin Bouniot, Benjamin M. Smith +13D Object Detection Medical Image Benchmarks
LTCI, Telecom Paris, Insitut Polytechnique de Paris, Palaiseau, France · Department of Medicine, Columbia University, New York, USA
Spatial Graph Representation and Morphometric Analysis of the Pulmonary Vascular Tree From Computed Tomography Using Multi-Scale Hessian-Based Filter Fusion and TEASAR Skeletonization
Jul 5, 2026Piotr Mackiewicz, Jakub Kołyska, Radoslaw Roszczyk3D Medical Imaging Medical Image Segmentation
Faculty of Electrical Engineering Warsaw University of Technology Warsaw, Poland
Anchored Slice Propagation with Boundary Transport for 3D CT Segmentation from Three Orthogonal Annotated Slices
Nov 3, 2026Tushar Kataria, Md Rakibul Haque, Shireen Y. ElhabianMedical Image Segmentation 3D Medical Image Segmentation
Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, USA · Kahlert School of Computing, University of Utah, Salt Lake City, USA

Apr 27, 2026J/K move · Enter open · S save

BifDet: A 3D Bifurcation Detection Dataset for Airway-Tree Modeling

Ali Keshavarzi, Quentin Bouniot, Benjamin M. Smith, Elsa Angelini

LTCI, Telecom Paris, Insitut Polytechnique de Paris, Palaiseau, France · Department of Medicine, Columbia University, New York, USA

3D Object Detection Medical Image Benchmarks

Open paper AskPDF arXiv

Thoracic Computed Tomography (CT) scans offer detailed insights into the intricate branching network of the airway tree, which is essential for understanding various respiratory diseases. Airway bifurcations, where airway branches split, are crucial landmarks for understanding lung physiology, disease mechanisms and lesion localization. Despite the significance of bifurcation analysis, a notable lack of datasets annotated for this task hinders the development of advanced automated specialized detection or segmentation tools. In this paper, we introduce BifDet, the first publicly-available dataset specialized for 3D airway bifurcation detection, filling a critical gap in existing resources. Our dataset comprises carefully annotated CT scans from the ATM22 open-access cohort with bifurcation bounding boxes covering the parent and daughter branches. As a use-case for demonstrating the potential of BifDet, we fine-tune and evaluate RetinaNet and DETR for 3D airway bifurcations detection on CT scans. We provide detailed pipelines, including preprocessing steps and specific implementation design choices. Results are detailed over various categories of minimal bounding box sizes to serve as baseline to benchmark future research.

Abstract

Image Segmentation

Explore similar work

BifDet: A 3D Bifurcation Detection Dataset for Airway-Tree Modeling

Spatial Graph Representation and Morphometric Analysis of the Pulmonary Vascular Tree From Computed Tomography Using Multi-Scale Hessian-Based Filter Fusion and TEASAR Skeletonization

Anchored Slice Propagation with Boundary Transport for 3D CT Segmentation from Three Orthogonal Annotated Slices