RABC-Net: Reliability-Aware Annotation-Free Skin Lesion Segmentation for Low-Resource Dermoscopy

Yujie Yao; Yuhaohang He; Junjie Huang; Zhou Liu; Jiangzhao Li; Yan Qiao; Wen Xiao; Yunsen Liang

cs.CVApr 7, 2026

RABC-Net: Reliability-Aware Annotation-Free Skin Lesion Segmentation for Low-Resource Dermoscopy

Authors: Yujie Yao, Yuhaohang He, Junjie Huang, Zhou Liu, Jiangzhao Li, Yan Qiao, Wen Xiao, Yunsen Liang, +1 more

Organizations: College of Information Engineering, Sichuan Agricultural University, Ya’an, China

Abstract

Pixel-level annotation is costly in low-resource dermoscopy. We present RABC-Net, a reliability-aware annotation-free segmentation system that combines pseudo-label reliability learning, restricted target-domain adaptation, and Reliability-Adaptive Boundary Calibration (RABC). The system decouples reliability learning from deployment: uncertainty-aware pseudo-label interaction shapes robust representations during training, while the image-only inference path is preserved and RABC performs local logit-space calibration from boundary confidence, uncertainty, and foreground probability. No manual masks are used for training or target-domain adaptation; validation labels, when available, are used only for final operating-point selection. Across ISIC-2017, ISIC-2018, and PH2, RABC-Net achieves macro-average DICE/JAC of 86.58%/79.47% and consistent matched-protocol results. Controlled within-study analyses show that RABC provides localized gains over nonlearned boundary correction, while the overall result comes from the full reliability-aware system. Adaptation updates only 3.50% of model parameters, image-only inference runs at 87.4 FPS, and the selected operating points use $σ=0$ on all three datasets, indicating that learned calibration avoids extra smoothing at deployment.

Figures & tables

Figure 1: Overall method overview. Four unsupervised paths produce a 4-channel pseudo-label prior tensor for IPC/PIA interaction. The image branch extracts multi-scale features, reliability heads estimate pseudo-label reliability, and the deployed image path uses RABC to calibrate logits before thresholding.

Path	Feature Source	Binarization Method
$P_{1}$	ResNet-50 layer1 spatial features	K-means ( $k=2$ ) + lesion cluster selection
$P_{2}$	ResNet-50 layer4 Grad-CAM heatmap	Otsu thresholding
$P_{3}$	ResNet-18 layer1 spatial features	K-means ( $k=2$ ) + lesion cluster selection
$P_{4}$	ViT-B/16 patch token features	K-means ( $k=2$ ) + lesion cluster selection

Table 1: Pseudo-label path definitions and feature sources.

Figure 2: Uncertainty-guided IPC/PIA correction. Image features and pseudo-label features interact through IPC and PIA attention, sigma heads produce reliability-aware weights, and the resulting weighted consensus probability map is binarized into a cleaner consensus mask for supervision.

Figure 3: Training-stage view of RABC-Net. The figure shows the shared forward path, reliability-weighted pseudo supervision, restricted branch adaptation, and the loss groups used for annotation-free optimization.

Figure 4: Inference-stage operating-point pipeline. RABC-calibrated logits are converted to a probability map, optional TTA and reference smoothing are applied only when selected, and lightweight morphology yields the final mask.

Setting	Stage 1	Stage 2
Training split	ISIC-2017	ISIC-2018 / PH2
Trainable modules	Full model	pseudo_encoder + IPC/PIA + $\sigma$ heads
Frozen deployed image-only path	No	Yes
Learning rate	$6\times 10^{-6}$	$5\times 10^{-6}$
Batch size	16	12
Epochs	24	2

Table 2: Key differences between Stage 1 source training and Stage 2 interaction-branch adaptation.

Metric	Value
Total parameters	31.47 M
ConvNeXt-Tiny backbone parameters	28.59 M
Decoder/head additional parameters	2.83 M (9.9% of backbone)
GFLOPs (image inference only)	57.39
GFLOPs (with pseudo-label interaction)	59.59
GFLOPs (ConvNeXt-Tiny backbone only)	8.91

Table 3: Model complexity and runtime characteristics of the proposed framework. The highlighted rows summarize the main deployment takeaway: lightweight image-only inference, low memory use, and restricted adaptation overhead.

Dataset	$\tau$	$\sigma$	TTA	fill_holes	keep_largest
ISIC-2017	0.30	0.0	flip4	$\checkmark$	$\checkmark$
ISIC-2018	0.25	0.0	flip4	$\checkmark$	$\checkmark$
PH2	0.06	0.0	flip4	$\checkmark$	$\checkmark$

Table 4: Validation-selected operating points used for final test evaluation.

Method	ACC	DICE	JAC	SEN	SPE
Sp. Merging	84.53	70.11	68.89	69.82	92.79
DRC	83.85	68.71	56.05	70.00	97.00
Saliency-CCE	85.51	72.85	61.91	77.44	94.34
A2S-v2	86.22	75.01	65.94	72.13	97.35
SpecWRSC	81.20	68.97	58.41	69.39	86.34
NCut	82.48	69.11	58.78	68.12	88.31

Table 5: Comparison with unsupervised methods on ISIC-2018.

Method	ACC	DICE	JAC	SEN	SPE
Sp. Merging	79.89	54.66	46.02	59.16	88.72
DRC	83.77	59.11	45.40	70.35	95.58
Saliency-CCE	83.87	61.77	49.53	74.07	92.98
A2S-v2	82.85	61.35	51.10	68.65	92.80
SpecWRSC	83.03	61.01	50.52	68.21	89.96
NCut	83.79	62.13	51.84	67.12	90.90

Table 6: Comparison with unsupervised methods on ISIC-2017.

Method	ACC	DICE	JAC	SEN	SPE
Sp. Merging	89.10	82.83	73.97	79.63	96.00
DRC	82.63	72.41	59.76	69.18	97.65
Saliency-CCE	84.62	76.79	65.62	78.90	93.77
A2S-v2	85.66	78.21	67.15	87.56	91.37
SpecWRSC	85.96	78.12	68.54	79.08	89.74
NCut	88.15	80.17	70.51	76.16	94.94

Table 7: Comparison with unsupervised methods on PH2.

Method	ACC	DICE	JAC	SEN	SPE
Reproduced RPI-Net (same-pipeline)	88.75	77.36	71.10	77.36	92.72
Ours	91.93	85.67	79.14	83.67	97.85

Table 8: Same-pipeline anchor comparison on ISIC-2018 under the shared validation-selected inference search space.

Figure 5: Qualitative segmentation comparison. Columns show image, GT, raw P0, final prediction, and error map (green=TP, red=FP, yellow=FN).

Figure 6: Qualitative comparison on ISIC-2018. Columns show image, GT, reproduced RPI-Net anchor, prompted MedSAM reference, final prediction, and final-model error map.

ID	Configuration	Thr	ACC	DICE	JAC	SEN	SPE
A0	Baseline (ConvNeXt-Tiny + IPC/PIA + DUL)	0.65	87.76	76.37	67.57	76.04	94.26
A1	A0 + Source distillation	0.50	90.30	82.15	74.20	79.66	97.01
A2	A1 + Stage 2 interaction-branch adaptation	0.37	91.40	85.08	77.83	83.01	98.02

Table 9: Training-side ablation (A0–A2) on ISIC-2018. A2 includes Stage 2 interaction-branch adaptation together with its own validation-selected operating point.

Figure 7: Pseudo-label refinement example. IPC/PIA converts divergent pseudo-labels into a reliability-weighted consensus before final prediction.

ID	Configuration	$\tau$	$\sigma$	TTA	ACC	DICE	JAC	SEN	SPE	15/15
P0	Raw	0.30	0.0	none	90.88	78.81	68.44	73.38	98.32	$\times$
P1	+ Threshold tuning	0.02	0.0	none	90.61	77.64	67.12	87.43	92.77	$\times$
P2	+ fill-holes + keep-largest	0.02	0.0	none	91.79	80.73	71.88	84.74	94.52	$\checkmark$
P3	+ reference smoothing ( $\sigma$ =1.0)	0.02	1.0	none	91.79	80.78	71.73	85.69	94.36	$\checkmark$
P4	+ TTA flip4	0.02	1.0	flip4	91.83	81.03	72.05	86.45	94.08	$\checkmark$

Table 10: Inference-side ablation pipeline on ISIC-2017, including optional reference smoothing.

ID	Configuration	$\tau$	$\sigma$	TTA	ACC	DICE	JAC	SEN	SPE	15/15
P0	Raw	0.30	0.0	none	91.47	85.16	77.91	84.04	97.56	$\times$
P1	+ Threshold tuning	0.13	0.0	none	91.13	84.22	76.50	87.28	94.96	$\checkmark$

Table 11: Inference-side ablation pipeline on ISIC-2018, including optional reference smoothing.

ID	Configuration	$\tau$	$\sigma$	TTA	ACC	DICE	JAC	SEN	SPE	15/15
P0	Raw	0.30	0.0	none	94.18	91.63	85.46	89.70	98.23	$\times$
P1	+ Threshold tuning	0.10	0.0	none	92.98	86.94	77.88	93.17	92.66	$\times$
P2	+ fill-holes + keep-largest	0.10	0.0	none	94.71	92.15	86.38	93.03	95.55	$\times$
P3	+ reference smoothing ( $\sigma$ =1.0)	0.10	1.0	none	94.69	91.98	86.09	93.57	95.26	$\times$
P4	+ TTA flip4	0.10	1.0	flip4	94.77	92.14	86.35	93.78	95.09	$\checkmark$

Table 12: Inference-side ablation pipeline on PH2, including optional reference smoothing.

Figure 8: SEN–SPE frontiers for optional reference smoothing on the no-RABC backbone. RABC is evaluated separately as learned calibration.

Method	DICE	JAC	SEN	SPE
Base	85.66	79.14	84.20	97.49
Threshold shift ( $\tau=0.13$ )	84.76	78.31	84.83	96.18
Dilation ( $1$ iter)	85.30	78.50	85.42	96.87
RABC	85.67	79.16	84.25	97.46

Table 13: Fixed-protocol ISIC-2018 control. The selected no-RABC protocol is fixed first, then threshold shift and dilation are matched to the validation SEN of RABC instead of being tuned for overlap.

Dataset	Method	DICE	JAC	SEN	SPE	HD95 $\downarrow$	ASSD $\downarrow$
ISIC-2017	Base	81.83	72.75	81.34	97.27	20.49	9.13
ISIC-2017	Dil.	81.67	72.55	81.68	97.35	20.44	9.16
ISIC-2017	RABC	81.90	72.86	81.57	97.21	20.44	9.09
ISIC-2018	Base	85.66	79.14	84.20	97.49	21.29	8.96
ISIC-2018	Dil.	85.01	78.16	84.83	97.08	21.42	9.30
ISIC-2018	RABC	85.68	79.15	83.68	97.85	20.93	8.84

Table 14: Controlled within-study comparison of Base, validation-selected binary dilation (Dil.), and RABC under the same evaluation pipeline. Lower is better for HD95/ASSD.

Configuration	ACC	DICE	JAC	SEN	SPE
Full (all modules enabled)	90.88	78.81	68.44	73.38	98.32
- Context Block (CSCB)	86.64	68.32	57.59	78.17	90.80
- Detail Block (DAB)	87.08	69.06	58.07	75.92	91.46
- Boundary Block (BRB)	86.78	68.69	57.97	77.13	90.98
- All Decoder Modules	87.03	70.11	59.14	80.14	89.56

Table 15: Decoder module-group ablation on ISIC-2017.

Figure 9: Ablation visualization for training strategy and decoder module groups.

Target dataset	Method	Protocol	DICE	JAC	SEN	SPE
ISIC-2018	no-RABC	Raw-P0 fixed	85.16	77.91	84.04	97.56
ISIC-2018	no-RABC	Source-transferred shared	83.97	77.41	85.70	94.52
ISIC-2018	RABC	Raw-P0 fixed	85.17	77.92	84.09	97.53
ISIC-2018	RABC	Source-transferred shared	85.51	78.94	82.90	98.23
PH2	no-RABC	Raw-P0 fixed	91.63	85.46	89.70	98.23
PH2	no-RABC	Source-transferred shared	92.28	86.59	93.90	94.96

Table 16: Target-domain controls without target-domain validation labels. Raw-P0 uses

\tau=0.30

without TTA or morphology. Source-transferred protocols select one operating point on ISIC-2017 validation and reuse it on both targets.

Target dataset	Method	DICE	JAC	SEN	SPE
ISIC-2018	no-RABC	83.97	77.41	85.71	94.52
ISIC-2018	Dil.	83.25	76.18	86.50	93.66
ISIC-2018	RABC	85.51	78.94	82.90	98.23
PH2	no-RABC	92.28	86.59	93.90	94.96
PH2	Dil.	91.76	85.70	94.97	94.34
PH2	RABC	91.68	85.60	89.57	98.03

Table 17: Target-free triplet comparison using operating points selected only on ISIC-2017 validation data.

Dataset	no-RABC			RABC
	ECE	Brier	NLL	ECE	Brier	NLL
ISIC-2017	0.3135	0.3142	1.1256	0.3083	0.3095	1.1025
ISIC-2018	0.2908	0.2847	1.0877	0.2902	0.2841	1.0829
PH2	0.2953	0.2945	1.0225	0.2951	0.2943	1.0209

Table 18: Boundary-band probability calibration before thresholding or morphology; lower is better.

Figure 10: Per-sample metric distributions on the three test sets.

Dataset	Method	DICE	JAC	HD95	ASSD
ISIC-2017	Reproduced RPI-Net	68.83	60.06	38.02	19.26
ISIC-2017	Ours	81.90	72.86	20.44	9.09
ISIC-2018	Reproduced RPI-Net	77.38	71.14	34.33	15.88
ISIC-2018	Ours	85.67	79.14	20.93	8.84
PH2	Reproduced RPI-Net	85.77	80.02	22.30	10.45
PH2	Ours	92.17	86.41	11.31	4.53

Table 19: Boundary distance comparison against a reproduced RPI-Net baseline.

Metric	RPI-Net	Ours
Parameters (M)	105.12	31.47
GPU Memory (MB)	484.73	686

Table 20: Model efficiency comparison between RPI-Net and the proposed method.

Dataset	Supervised mean DICE	Ours (unsup.)	Gap
ISIC-2017	84.59	81.90	2.69 pp
ISIC-2018	88.01	85.67	2.34 pp
PH2	93.16	92.17	0.99 pp

Table 21: Gap between the proposed annotation-free training framework and the mean DICE of six supervised reference models.

Method	ACC	DICE	JAC	SEN	SPE
U-Net	93.87	85.43	77.31	83.02	97.61
U-Net++	93.24	84.45	76.05	80.78	98.60
Att-DeepLabv3+	93.51	84.82	76.24	80.33	98.56
TransFuse	93.77	85.26	76.91	82.62	98.08
FAT-Net	93.41	84.97	76.51	81.06	98.51
MALUNet	92.32	82.59	73.57	82.12	97.79

Table 22: Comparison with supervised methods on ISIC-2017.

Method	ACC	DICE	JAC	SEN	SPE
U-Net	93.33	88.18	80.40	94.99	92.45
U-Net++	93.72	88.23	80.65	93.86	94.34
Att-DeepLabv3+	93.42	88.20	80.50	94.56	93.18
TransFuse	93.85	88.46	80.80	94.94	93.05
FAT-Net	93.73	88.95	81.87	92.75	94.92
MALUNet	92.07	86.04	77.87	90.89	94.57

Table 23: Comparison with supervised methods on ISIC-2018.

Method	ACC	DICE	JAC	SEN	SPE
U-Net	96.66	94.14	89.32	94.76	97.37
U-Net++	95.18	90.80	83.61	95.25	93.05
Att-DeepLabv3+	95.84	93.29	88.03	95.02	96.70
TransFuse	96.69	93.76	88.73	97.39	96.05
FAT-Net	96.49	93.63	88.93	96.39	96.03
MALUNet	96.45	93.32	87.89	96.07	96.26

Table 24: Comparison with supervised methods on PH2.

Explore similar work

MDSkin-Net: Multi-Task Skin Lesion Analysis Driven by Pattern Analysis Priors and Spatial Alignment Regularization
Sep 25, 2026Yijian Li, Saad Bedros, Paul Bigliardi +3Skin Lesion Classification Skin Lesion Segmentation
University of Minnesota, Minneapolis, MN, USA
Less Is More: A Leakage-Controlled Study of Dermoscopic Preprocessing for Joint Skin Lesion Classification and Segmentation with YOLO26
Oct 6, 2026Truong Viet Vu, Nguyen Chi Hai, Nguyen Phuc Nguyen +3Skin Lesion Classification Medical Image Segmentation
Faculty of Information Technology, Van Lang School of Technology, Van Lang University, Ho Chi Minh City, Vietnam
Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification
Jul 21, 2026Abu Mukaddim Rahi, Md Mithun Hossain, Md Zulficar Hasan Joy +2Skin Lesion Classification Medical Image Classification
Department of IT in Artificial Intelligence, Victoria University, Footscray, VIC, Australia · BUBT Resarch Graduate School, Bangladesh University of Business and Technology, Dhaka, Bangladesh · Department of Artificial Intelligence, Daffodil Institute of Information Technology, Dhaka, Bangladesh +2

Sep 25, 2026J/K move · Enter open · S save

MDSkin-Net: Multi-Task Skin Lesion Analysis Driven by Pattern Analysis Priors and Spatial Alignment Regularization

Yijian Li, Saad Bedros, Paul Bigliardi, Mei Bigliardi Qi+2

University of Minnesota, Minneapolis, MN, USA

Skin Lesion Classification Skin Lesion Segmentation Multi-Task Learning Domain Generalization

Open paper AskPDF arXiv

Reliable skin lesion segmentation and classification are central to dermoscopic computer-aided diagnosis. Existing multi-task frameworks couple the two tasks architecturally without clinical knowledge, while knowledge-injecting approaches rely on the macroscopic ABCD rule, which was not designed for dermoscopy. Dermoscopic diagnosis is grounded in Pattern Analysis, a microscopic framework structured around dermoscopic features. We propose MDSkin-Net, which incorporates cue-level Pattern Analysis priors into a hybrid CNN-Transformer architecture. At its core is a Pattern Analysis-Guided Attention Module (PAGAM) comprising three priors motivated by distinct dermoscopic cues: an improved Efficient Channel Attention (iECA), a Multi-Scale Spatial Attention (MSSA), and a Biased Asymmetry Attention (BAA). We further introduce a multi-scale spatial alignment regularization (MSAR) that uses the segmentation ground-truth mask as hierarchical soft supervision, confining the classification head to lesion-localized evidence and coupling both task pathways through a shared spatial prior. Trained exclusively on the ISIC 2017 training split without external dermoscopy data, the MDSkin-Net ensemble transfers robustly under zero-shot evaluation, reaching a Dice Similarity Coefficient (DSC) of 92.38% and a melanoma AUC of 97.84%on PH2, and a DSC of 88.92% on the ISIC 2018 Task 1 test set. On the in-domain ISIC 2017 benchmark, the ensemble attains a mean Area Under the Curve (AUC) of 91.60% across the two classification tasks (melanoma and seborrheic keratosis vs. rest), and a DSC of 84.72% for segmentation. Classification remains competitive with baselines; in-domain segmentation trails single-task specialists, yet the proposed priors and alignment regularization yield representations that generalize consistently across cohorts of different scales.

Abstract

Explore similar work

MDSkin-Net: Multi-Task Skin Lesion Analysis Driven by Pattern Analysis Priors and Spatial Alignment Regularization

Less Is More: A Leakage-Controlled Study of Dermoscopic Preprocessing for Joint Skin Lesion Classification and Segmentation with YOLO26

Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification