Domain-Guided YOLO26 with Composite BCE-Dice-Lovász Loss for Multi-Class Fetal Head Ultrasound Segmentation

M. Fazri Nizar

cs.CVMar 23, 2026

Domain-Guided YOLO26 with Composite BCE-Dice-Lovász Loss for Multi-Class Fetal Head Ultrasound Segmentation

Authors: M. Fazri Nizar

Organizations: Informatics Engineering, Faculty of Computer Science Sriwijaya University

Abstract

Segmenting fetal head structures from prenatal ultrasound remains a practical bottleneck in obstetric imaging. The current state-of-the-art baseline, proposed alongside the published dataset, adapts the Segment Anything Model with per-class Dice and Lovász losses but still depends on bounding-box prompts at test time. We build a prompt-free pipeline on top of YOLO26-Seg that jointly detects and segments three structures, Brain, Cavum Septi Pellucidi (CSP), and Lateral Ventricles (LV), in a single forward pass. Three modifications are central to our approach: (i) a composite BCE-Dice-Lovász segmentation loss with inverse-frequency class weighting, injected into the YOLO26 training loop via runtime monkey-patching; (ii) domain-guided copy-paste augmentation that transplants minority-class structures while respecting their anatomical location relative to the brain boundary; and (iii) inter-patient stratified splitting to prevent data leakage. On 575 held-out test images, the composite loss variant reaches a mean Dice coefficient of 0.9253, exceeding the baseline (0.9012) by 2.68 percentage points, despite reporting over three foreground classes only, whereas the baseline's reported mean includes the easy background class. We further ablate each component and discuss annotation-quality and class-imbalance effects on CSP and LV performance.

Figures & tables

Split	Images	Brain	CSP	LV	Total
Train	2,654	2,697	921	1,052	4,670
Val	603	611	190	212	1,013
Test	575	568	172	212	952
All	3,832	3,792	1,283	1,476	6,551

TABLE I: Instance counts after inter-patient splitting.

Fig. 1: Class distribution and imbalance ratio across the full dataset.

Fig. 2: Per-split class counts, showing that the stratified split preserves proportional representation.

Variant	Speed CPU ONNX (ms)	Speed GPU T4 TRT10 (ms)	Params	FLOPs
YOLO26N-Seg	53.3 $\pm$ 0.5	2.1 $\pm$ 0.0	2.7 M	9.1
YOLO26L-Seg	387.0 $\pm$ 3.7	8.0 $\pm$ 0.1	28.0 M	139.8

TABLE II: Model variants used in this work.

Class	Instances	Weight	Normalised
Brain	3,792 (57.9%)	0.1	0.059
CSP	1,283 (19.6%)	0.9	0.529
LV	1,476 (22.5%)	0.7	0.412

TABLE III: Inverse-frequency class weights used in Dice and Lovász terms.

Parameter	Baseline	Domain-Guided	FetSAM Loss
Architecture	YOLO26N-Seg	YOLO26L-Seg	YOLO26N-Seg
Params	2.7 M	28.0 M	2.7 M
Input size	640	800	640
Batch size	64	16	128
Epochs	180 (early stop)	300 (3 sess.)	300
Optimizer	AdamW	Auto $\to$ MuSGD	Auto $\to$ AdamW

TABLE IV: Hyperparameters per experiment.

Experiment	Sessions	Epochs	Total time	Per epoch
Baseline	1	180	2.45 h	48.8 s
Domain-Guided	3	300	30.88 h	368.3 s
FetSAM Loss	1	300	4.04 h	48.3 s

TABLE V: Wall-clock training times on T4

\times

Metric	Baseline	Domain-Guided	FetSAM Loss
mAP@50	0.7423	0.8657	0.7700
mAP@50-95	0.4736	0.5462	0.5046
mDSC	0.9229	0.9208	0.9253
mIoU	0.8661	0.8624	0.8697
Precision	0.7393	0.8244	0.7592
Recall	0.7810	0.8775	0.8178

TABLE VI: Test-set results (575 images). Best values in bold.

Fig. 3: Grouped bar chart of detection, segmentation, and classification metrics.

Class	Prec.	Rec.	F1	IoU	DSC	TP	FP	FN
Brain	0.939	0.982	0.960	0.943	0.971	558	36	10
CSP	0.624	0.762	0.686	0.704	0.823	131	79	41
LV	0.655	0.599	0.626	0.695	0.816	127	67	85

TABLE VII: Per-class results, Baseline (YOLO26N, standard BCEDice).

Class	Prec.	Rec.	F1	IoU	DSC	TP	FP	FN
Brain	0.984	0.997	0.990	0.947	0.973	566	9	2
CSP	0.700	0.895	0.786	0.722	0.836	154	66	18
LV	0.789	0.741	0.764	0.693	0.816	157	42	55

TABLE VIII: Per-class results, Domain-Guided (YOLO26L, BCE-Dice-Lovász + domain aug.).

Class	Prec.	Rec.	F1	IoU	DSC	TP	FP	FN
Brain	0.983	0.998	0.990	0.946	0.972	567	10	1
CSP	0.640	0.837	0.725	0.715	0.831	144	81	28
LV	0.655	0.618	0.636	0.712	0.828	131	69	81

TABLE IX: Per-class results, FetSAM Loss (YOLO26N, BCE-Dice-Lovász only).

Fig. 4: Per-class DSC across the three models, with numeric labels.

Fig. 5: Metric heatmap for the Domain-Guided model; warm colours highlight the Brain–CSP–LV performance disparity.

Fig. 6: Training curves: segmentation loss, box loss, classification loss, mAP@50, mAP@50-95, and validation segmentation loss versus epoch.

Fig. 7: Learning-rate schedules. All runs use cosine annealing with warm restarts [ 35 ] ; the Domain-Guided curve resets at session boundaries.

Model	Backbone	Paradigm	Mean DSC
FetSAM [ 10 ]	SAM (ViT)	Prompt-based	0.9012
Efb0_DLabV3+ [ 10 ]	EffNet-B0 [ 36 ]	Semantic	0.8532
Mit_SegFormer_b0 [ 10 ]	MiT-B0 [ 7 ]	Semantic	0.8323
Efb0_Unet [ 10 ]	EffNet-B0 [ 36 ]	Semantic	0.8466
Efb0_Unet++ [ 10 ]	EffNet-B0 [ 36 ]	Semantic	0.8495
Efb0_Linknet [ 10 ]	EffNet-B0 [ 36 ]	Semantic	0.8184

TABLE X: Mean DSC comparison with models reported in Alzubaidi et al. [ 10 ] .

Class	FetSAM	Ours (FetSAM Loss)	Ours (Dom.-Guided)
Brain	0.9851	0.9718	0.9729
CSP	0.8037	0.8305 ( $+$ 3.3%)	0.8359 ( $+$ 4.0%)
LV	0.8208	0.8280 ( $+$ 0.9%)	0.8162
Mean (fg)	0.9012*	0.9253	0.9208
*FetSAM’s published mean includes the background class.

TABLE XI: Per-class DSC comparison with FetSAM [ 10 ] .

Fig. 8: Vertex counts per annotation. Brain contours (often $>$ 50 vertices) are substantially more detailed than CSP and LV masks (frequently 4–8 vertices).

Fig. 9: Model predictions (contours) versus ground truth (filled) on test images containing all three classes. The model often produces smoother boundaries than the coarsely annotated references.

Ultrasound Image Segmentation

Ultrasound image segmentation delineates organs, lesions, and other structures in ultrasound scans for measurement, diagnosis, and image-guided care.

See all 73 papers

FETALFusion: Resolution-Aware State-Space Modeling for Multi-Domain Fetal Head Segmentation
Oct 13, 2026Ammar Bhilwarawala, Mainak BandyopadhyayMedical Image Segmentation Ultrasound Image Segmentation
School of Computer Engineering, KIIT Deemed to be University Bhubaneswar, India
Dual Agreement Consistency Learning for Semi-Supervised Fetal Ultrasound Segmentation
Jun 24, 2026Fangyijie Wang, Guénolé Silvestre, Ziyang Wang +1Ultrasound Image Segmentation Semi-Supervised Medical Image Segmentation
Research Ireland Centre for Research Training in Machine Learning · School of Medicine, University College Dublin, Dublin, Ireland · School of Computer Science, University College Dublin, Dublin, Ireland +1
Attention-ResUNet for Automated Fetal Head Segmentation
Apr 20, 2026Ammar Bhilwarawala, Mainak BandyopadhyayUltrasound Image Segmentation Medical Image Segmentation
School of Computer Engineering, KIIT Deemed to be University, Bhubaneswar, India

Oct 13, 2026J/K move · Enter open · S save

FETALFusion: Resolution-Aware State-Space Modeling for Multi-Domain Fetal Head Segmentation

Ammar Bhilwarawala, Mainak Bandyopadhyay

School of Computer Engineering, KIIT Deemed to be University Bhubaneswar, India

Medical Image Segmentation Ultrasound Image Segmentation Image Segmentation Visual SSMs

Open paper AskPDF arXiv

Head circumference (HC) measured from 2D ultrasound is a routine marker of fetal growth, and automating it needs segmentations whose boundaries hold up on individual scans, not only on average. We present FETALFusion, a dual-path CNN-Mamba encoder-decoder for fetal head segmentation across acquisition types. It applies four-directional state-space scans only at the two highest-resolution stages, routes bottleneck features through domain-specific channel gates that are averaged at test time, and adds a weak landmark shape prior. We train jointly on transabdominal (HC-18) and transperineal (PSFHS) images and compare ten methods on a held-out test set of 471 images, each trained under one schedule and scored by surface distances at native resolution. FETALFusion reaches a Dice score of 0.963 and an HD95 of 2.10 mm without test-time augmentation or a domain label. It improves on U-Mamba on all five reported metrics (HD95 reduced by 0.31 mm; Holm-corrected (

p < 10^{-4}

), a lead that holds under test-time augmentation and across two independent training runs, and in the mean on U-Net, Attention U-Net and TransUNet. Residual encoders close the gap: a ResNet-34 UNet++ exceeds FETALFusion, and an ImageNet initialised ResNet-34 U-Net is best overall (HD95 1.75 mm). The ablation agrees. Calibrated against run-to-run variation, it attributes FETALFusion's measurable gains to the landmark shape prior and, on transperineal images, to the domain gates, while removing the state-space branch does not measurably change accuracy. HC from the predicted masks agrees with the clinical reference at an intraclass correlation of 0.9986.

Abstract

Ultrasound Image Segmentation

Explore similar work

FETALFusion: Resolution-Aware State-Space Modeling for Multi-Domain Fetal Head Segmentation

Dual Agreement Consistency Learning for Semi-Supervised Fetal Ultrasound Segmentation

Attention-ResUNet for Automated Fetal Head Segmentation