CLC-YOLO: A Compact Channel-Gated Prototype Network for Real-Time Leakage-Aware Breast Ultrasound Lesion Segmentation

M. Fazri Nizar; Muhammad Naufal Rachmatullah; Julian Supardi

cs.CVSep 19, 2026

CLC-YOLO: A Compact Channel-Gated Prototype Network for Real-Time Leakage-Aware Breast Ultrasound Lesion Segmentation

Authors: M. Fazri Nizar, Muhammad Naufal Rachmatullah, Julian Supardi

Organizations: Department of Informatics Engineering, Universitas Sriwijaya, Palembang, South Sumatera

Abstract

Reliable breast ultrasound lesion segmentation requires accurate boundaries and evaluation that prevents patients or duplicate images from crossing data splits. We propose Channel Local Contrast (CLC), a compact refinement of the YOLO26 segmentation prototype head. CLC adds a fixed local high-pass residual controlled by 64 zero-initialized, bounded channel gates. Baseline and CLC were compared in five matched folds on each of four breast ultrasound datasets. BUS-BRA used patient-disjoint outer tests with separate inner validation. BUS-UCLM and BrEaST used patient-grouped validation folds; BUSI used duplicate-component groups because patient identifiers are unavailable. Group-macro Dice increased by 1.68, 3.11, 1.12, and 2.45 percentage points on BUS-BRA, BUS-UCLM, BUSI, and BrEaST, respectively. Only the BUS-BRA paired 95% confidence interval excluded zero. CLC adds 64 parameters and 0.0049 giga floating-point operations (GFLOPs). At 640 pixels, single-T4, batch-one, 16-bit floating-point (FP16) TensorRT graph times were 2.1478 ms for CLC and 2.0512 ms for baseline, excluding preprocessing and postprocessing. CLC increased group-macro Dice across all four datasets with a measured T4 forward-pass overhead of 0.0966 ms. Code: https://github.com/mfazrinizar/CLC-YOLO

Figures & tables

Dataset	Images	Lesion / Normal	Evaluation Group	Diagnostic Reference	Segmentation Reference
BUS-BRA [ 1 ]	1,875	1,875 / 0	Inter-patient: 1,064 patient IDs, official Case folds	Biopsy-proven diagnosis (histopathology not specified)	Senior ultrasonographer mask
BUS-UCLM [ 2 ]	683	264 / 419	Inter-patient: 38 patient prefixes	Biopsy-confirmed lesion diagnosis (histopathology not specified)	Radiologist contour, second review
BUSI [ 3 ]	761	630 / 131	Duplicate-component proxy: 579 groups, no patient IDs	Supplied class, histology not established	Supplied manual mask
BrEaST [ 4 ]	256	252 / 4	Inter-patient: 256 case IDs	Histopathologic diagnosis after biopsy (197) or >2-y follow-up (55)	Radiologist freehand mask

TABLE I: Dataset and medical-evaluation integrity.

Fig. 1: Dataset overview: raw display crop, ground-truth (GT) instance mask and box, and annotated crop. Inference uses full images.

Fig. 2: CLC-YOLO at nano scale. Backbone and PAN-FPN neck feed Segment26; green identifies the proposed prototype refinement. Black arrows follow sequential nodes, blue arrows route lateral tensors, and breaks denote crossings without a junction. Expanded in Fig. 3.

Fig. 3: CLC-Proto26: 64 bounded gates modulate local residuals before the unchanged prototype generator. Black/teal paths are inference; orange dashed paths are training-only. Mask coefficients combine 32 prototypes, followed by resize, threshold, and predicted-box cropping.

Variant	Added	BUS-BRA	BUS-UCLM	BUSI	BrEaST	Mean	Positive
Baseline	0	87.53	55.81	69.56	67.33	70.06	Reference
Global LC	1	88.86	52.79	67.03	63.22	67.97	1/4
Channel LC	64	89.21	58.93	70.69	69.78	72.15	4/4
HLC	66	87.68	52.99	69.07	65.47	68.80	1/4
DGC-LC	65	88.14	59.98	69.03	61.00	69.54	2/4
C3-LC	129	88.82	61.32	67.17	66.40	70.93	2/4

TABLE II: Five-fold architecture ablation using patient/proxy-group macro Dice (%). BUS-BRA is outer-test evaluation; the other cohorts are grouped validation.

Dataset	Dice	IoU	mAP50	mAP50-95	Paired Dice 95% CI	Fold Wins
BUS-BRA	89.21 (+1.68)	82.90 (+1.76)	85.41 (+1.33)	60.28 (+1.74)	[+0.76, +2.63]	4/5
BUS-UCLM	58.93 (+3.11)	53.27 (+3.13)	71.61 (+0.91)	48.92 (+1.07)	[-2.70, +9.23]	3/5
BUSI	70.69 (+1.12)	64.30 (+0.85)	72.28 (+1.45)	47.78 (+0.80)	[-1.43, +3.59]	3/5
BrEaST	69.78 (+2.45)	62.67 (+2.54)	70.64 (+1.57)	44.40 (+1.93)	[-2.43, +7.43]	4/5

TABLE III: Paired CLC-YOLO scores (%), with gain over baseline in parentheses (percentage points). Gains use unrounded fold aggregates. The CI and fold wins apply to Dice.

Model	Train Parameters	GFLOPs	Ryzen CPU ONNX (ms)	T4 TensorRT 10 (ms)	Total Training (h)	Weighted s/Epoch
Baseline	3,053,510	8.957952	33.2815 ± 0.7092	2.0512 ± 0.0102	4.3861 (1,901 epochs)	8.3062
CLC-YOLO	3,053,574	8.962867	34.8376 ± 1.0644	2.1478 ± 0.0098	4.5104 (1,964 epochs)	8.2675
Difference	+64 (+0.0021%)	+0.004915 (+0.0549%)	+1.5561 (+4.68%)	+0.0966 (+4.71%)	+0.1243 (+2.83%)	-0.0387 (-0.47%)

TABLE IV: Complexity, inference latency, and training efficiency.

Fig. 4: Two benign and two malignant cases selected for favorable CLC Dice differences, with nonempty baseline masks. Columns: raw image, GT, baseline, CLC. Selection favors visible improvements over typical performance.

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

Specificity- and Calibration-Aware Breast Ultrasound Segmentation via Entropy-Guided Boundary Supervision
Jun 21, 2026Manar Alsaid, Mandip Shrestha, Mohammad AbbasUltrasound Image Segmentation Model Calibration
Department of Computer Science East Texas A&M University Commerce, USA · Texas Woman’s University Denton, USA
Adaptive Bidirectional Task Interaction for Joint Segmentation and Classification of Breast Ultrasound
Mar 1, 2026Abdullah Al Shafi, Md Kawsar Mahmud Khan Zunayed, Safin Ahmmed +2Ultrasound Image Segmentation Image Segmentation
Khulna University of Engineering and Technology, Bangladesh · University Clermont Auvergne, France
Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation
Jul 23, 2026Jingguo Qu, Xinyang Han, Xiang Wang +8Ultrasound Image Segmentation Implicit Neural Representations
Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China

Jun 21, 2026J/K move · Enter open · S save

Specificity- and Calibration-Aware Breast Ultrasound Segmentation via Entropy-Guided Boundary Supervision

Manar Alsaid, Mandip Shrestha, Mohammad Abbas

Department of Computer Science East Texas A&M University Commerce, USA · Texas Woman’s University Denton, USA

Ultrasound Image Segmentation Model Calibration Boundary-Aware Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Lesion segmentation in breast ultrasound involves two related challenges. In images with lesions, speckle noise, low tissue contrast, and posterior acoustic shadowing cause boundary leakage and incomplete contour delineation. In images without lesions, those same artifacts generate false-positive activations in regions resembling solid lesion tissue. This study addresses both failure modes through a single modification to the training objective. Rather than weighting every boundary pixel equally, the proposed loss scales contour penalties by per-pixel predictive entropy and the ground-truth boundary map, concentrating gradient emphasis on lesion margin locations where the network remains uncertain. The loss was evaluated on the BUSI dataset through a controlled ablation against two baselines: a model without boundary supervision and a model with uniformly weighted boundary binary cross-entropy. Across 97 lesion-containing test images, mean Dice scores were statistically indistinguishable between the proposed method and the no-boundary baseline (0.7624 versus 0.7616, paired Wilcoxon p = 0.27), confirming that lesion segmentation quality is preserved. The primary effect appears in specificity. False-positive activations on 20 no-lesion test images fell from 14 of 20 and 19 of 20 for the two baselines to 5 of 20 with the proposed approach (McNemar p = 0.012 and 0.0005). Non-overlapping Wilson 95% confidence intervals confirm the difference is both statistically significant and practically substantial. A post-hoc spatial temperature scaling step further reduced expected calibration error from 0.0201 to 0.0095 without altering segmentation masks. Entropy-guided boundary supervision and spatial calibration thus function as complementary training-level and inference-level refinements that improve specificity and probability reliability within a U-Net framework.

Abstract

Ultrasound Image Segmentation

Explore similar work

Specificity- and Calibration-Aware Breast Ultrasound Segmentation via Entropy-Guided Boundary Supervision

Adaptive Bidirectional Task Interaction for Joint Segmentation and Classification of Breast Ultrasound

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation