SAM3-SemiSkin: Parameter-Efficient Adaptation with Heterogeneous Decoders for Semi-Supervised Dermoscopic and Brain-MRI Segmentation

Background/Objectives: Reliable pixel-level segmentation is needed when lesion boundaries are faint, irregular or affected by acquisition artefacts, yet dense annotation is costly. We investigated whether heterogeneous decoders remain complementary when they share a frozen foundation-model representation, and separated this architectural effect from the contribution of unlabelled training. Methods: SAM3-SemiSkin combines a frozen SAM3 image encoder, bottleneck adapters, and convolutional neural network (CNN) and Transformer decoders. Agreement-weighted soft targets provide mutual supervision, supplemented by frequency and morphological objectives; colour compactness is restricted to dermoscopy. Evaluation included ISIC-2018 benchmarks, controls that retain the shared frozen SAM3 encoder and identical adapter stack, component analyses, zero-shot prompt controls, robustness perturbations, and two-dimensional TCGA-LGG brain magnetic resonance imaging (MRI). MRI data were partitioned by patient and evaluated slice-wise. Training and validation annotation budgets were counted separately. Results: On ISIC-2018, three-seed test means were 0.886/0.811 for mDice/mIoU at nominal 10% training labels and 0.894/0.825 at 30%. In a separate seed-42 checkpoint audit, the epochs with maximal validation mDice were 45 and 40; train/validation/test mDice were 0.964/0.929/0.882 and 0.974/0.942/0.891, respectively. The 30%-label complete ablation reached 0.893/0.827, exceeding a supervised single-CNN reference by 4.1 percentage points (pp) in mDice. Three-seed paired contrasts were 1.53±0.42 pp for heterogeneous versus homogeneous decoders and 0.63±0.38 pp for adding unlabelled objectives. Matched controls and zero-shot prompt evaluations contextualize these gains. At 30% labels, the MRI benchmark yielded mDice 0.833±0.007 and mIoU 0.740±0.007. Approximately 2.2% of the complete model parameters were trainable. Conclusions: The results support parameter-efficient adaptation with heterogeneous decoding as an integrated design for the tested dermoscopic and MRI tasks. The smaller unlabelled-training contrast and competitive matched controls limit the contribution attributable to the particular pseudo-label objective.

Authors

Institutions

Publication Details

Journal
Diagnostics
Published
2026-11-17
DOI
https://doi.org/10.3390/diagnostics16223837
Primary Topic
Medical Image Segmentation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

SAM3-SemiSkin: Parameter-Efficient Adaptation with Heterogeneous Decoders for Semi-Supervised Dermoscopic and Brain-MRI Segmentation

Gabriel Villarrubia González, Yanglin Du
4 links
Diagnostics
Medical Image Segmentation Techniques
article

SAM3-SemiSkin: Parameter-Efficient Adaptation with Heterogeneous Decoders for Semi-Supervised Dermoscopic and Brain-MRI Segmentation

Gabriel Villarrubia González, Yanglin Du
article en

Abstract

Background/Objectives: Reliable pixel-level segmentation is needed when lesion boundaries are faint, irregular or affected by acquisition artefacts, yet dense annotation is costly. We investigated whether heterogeneous decoders remain complementary when they share a frozen foundation-model representation, and separated this architectural effect from the contribution of unlabelled training. Methods: SAM3-SemiSkin combines a frozen SAM3 image encoder, bottleneck adapters, and convolutional neural network (CNN) and Transformer decoders. Agreement-weighted soft targets provide mutual supervision, supplemented by frequency and morphological objectives; colour compactness is restricted to dermoscopy. Evaluation included ISIC-2018 benchmarks, controls that retain the shared frozen SAM3 encoder and identical adapter stack, component analyses, zero-shot prompt controls, robustness perturbations, and two-dimensional TCGA-LGG brain magnetic resonance imaging (MRI). MRI data were partitioned by patient and evaluated slice-wise. Training and validation annotation budgets were counted separately. Results: On ISIC-2018, three-seed test means were 0.886/0.811 for mDice/mIoU at nominal 10% training labels and 0.894/0.825 at 30%. In a separate seed-42 checkpoint audit, the epochs with maximal validation mDice were 45 and 40; train/validation/test mDice were 0.964/0.929/0.882 and 0.974/0.942/0.891, respectively. The 30%-label complete ablation reached 0.893/0.827, exceeding a supervised single-CNN reference by 4.1 percentage points (pp) in mDice. Three-seed paired contrasts were 1.53±0.42 pp for heterogeneous versus homogeneous decoders and 0.63±0.38 pp for adding unlabelled objectives. Matched controls and zero-shot prompt evaluations contextualize these gains. At 30% labels, the MRI benchmark yielded mDice 0.833±0.007 and mIoU 0.740±0.007. Approximately 2.2% of the complete model parameters were trainable. Conclusions: The results support parameter-efficient adaptation with heterogeneous decoding as an integrated design for the tested dermoscopic and MRI tasks. The smaller unlabelled-training contrast and competitive matched controls limit the contribution attributable to the particular pseudo-label objective.

DiagnosticsVol. 16(22)
Universidad de Salamanca (ES)
Medical Image Segmentation Techniques
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