CG-MedSAM: Contrast-Gated Parameter-Efficient MedSAM for Skin-Tone-Robust Lesion Segmentation

Official PyTorch Weights & Model Card
MICCAI Submission — Paper ID: 1042
GitHub: raju-sah/MedSAM-For-Skin-Segmentation

---

Model Summary

CG-MedSAM is a parameter-efficient vision adaptation framework that introduces zero-leakage, contrast-gated residual bottleneck adapters into the frozen Vision Transformer (ViT-B) backbone of MedSAM.

Medical foundation vision models exhibit strong general-purpose segmentation capabilities, but suffer severe performance degradation when deployed across diverse skin-tone populations (Fitzpatrick Skin Types V–VI) and under imperfect clinical prompt conditions. In pigmented lesions on melanin-rich skin, optical contrast with surrounding healthy tissue is subtle, causing standard decoders to erode boundaries or leak into healthy skin.

CG-MedSAM resolves this by dynamically modulating adapter activations using a localized color-contrast proxy ($\Delta E^*_{ab}$) extracted strictly in the perceptually uniform CIE $L^a^b^*$ color space between an eroded lesion core ($C_{\text{core}}$) and an expanded perilesional background ring ($R_{\text{ring}}$). By conditioning adapter capacity on optical contrast difficulty with zero ground-truth label leakage, CG-MedSAM achieves superior boundary conformance on dark skin tones while updating only 4.64% of backbone parameters.

---

Architectural Highlights

• Base Backbone: Pre-trained MedSAM ViT-B ($89.67\text{M}$ frozen parameters).
• PEFT Adaptation: Parallel bottleneck adapters inserted into all 12 self-attention blocks with bottleneck rank $r=16$.
• Trainable Parameters: $4,363,824$ parameters ($4.64%$ of backbone weights).
• Dynamic Gating Mechanism: Lightweight Multi-Layer Perceptron: $$\gamma = g(\hat{c}) = 2.0 \cdot \text{Sigmoid}\left(\mathbf{W}_2 \cdot \text{ReLU}(\mathbf{W}_1 \hat{c} + \mathbf{b}_1) + b_2\right) \in (0, 2)$$
• Zero-Leakage Guarantee: Contrast features $\hat{c} = (\Delta E^*_{ab} - 28.5) / 12.0$ are derived strictly from the prompt bounding box and RGB input without requiring ground-truth annotation.

---

Quantitative Empirical Results

1. In-Domain Dermoscopy (ISIC 2018 Validation, $N=519$)

Architecture	Trainable Params	Param %	Clean ($\delta=0$)	5% Jitter	10% Jitter	20% Jitter	Degradation $\Delta\downarrow$
Zero-Shot MedSAM	0	0.00%	0.9302	0.9186	0.8857	0.7969	-0.1333
Decoder-Only	4,058,340	4.33%	0.9524	0.9460	0.9397	0.8980	-0.0544
LoRA ($r=16$)	4,648,164	4.93%	0.9609	0.9559	0.9514	0.9282	-0.0327
Standard Adapter ($r=16$)	4,362,660	4.64%	0.9633	0.9599	0.9536	0.9296	-0.0337
CG-Adapter (Ours)	4,363,824	4.64%	0.9638	0.9608	0.9553	0.9323	-0.0315

2. External Clinical Fairness Benchmark (sDDI Test, $N=198$, 9,900 evaluations)

Architecture	Clean Overall	Clean Light	Clean Med	Clean Dark	Clean Disparity $\Delta_{\text{L-D}}\downarrow$	20% Jitter Dark	Disparity $\Delta_{\text{L-D}}\downarrow$
Zero-Shot MedSAM	0.7694	0.7886	0.7671	0.7534	0.0351	0.6725	0.0500
Standard Adapter	0.8268	0.8336	0.8398	0.8022	0.0314	0.7650	0.0327
Decoder-Only	0.8415	0.8441	0.8480	0.8301	0.0139	0.7830	0.0147
CG-Adapter (Ours)	0.8402	0.8505	0.8449	0.8234	0.0270	0.7710	0.0341
LoRA ($r=16$)	0.8619	0.8651	0.8662	0.8530	0.0122	0.7789	0.0295

3. Statistical Significance

Paired two-sided Wilcoxon signed-rank hypothesis tests with Holm-Bonferroni correction ($\alpha=0.05$):

• CG-Adapter vs Standard Adapter (Clean): $\Delta = +0.0134$, $p_{\text{HB}} = \mathbf{6.11 \times 10^{-4}}$ ($p < 0.001$).
• CG-Adapter vs Standard Adapter (20% Jitter): $\Delta = +0.0061$, $p_{\text{HB}} = \mathbf{0.0318}$ ($p < 0.05$).
• CG-Adapter vs Zero-Shot MedSAM: $\Delta = +0.0707$, $p_{\text{HB}} = \mathbf{1.07 \times 10^{-13}}$, effect size $r_{\text{rb}} = +0.630$.

---

Intended Clinical Use & Out-of-Scope Restrictions

Intended Use

• Research investigations in algorithmic fairness and dermatological AI equity.
• Pre-segmentation assistance for computer-aided triage and dermoscopic feature extraction.
• Cross-population benchmarking across diverse Fitzpatrick skin types (FST I through VI).

Out-of-Scope / Clinical Restrictions

• Not for standalone medical diagnosis: Model outputs must always be reviewed by qualified dermatologists or trained clinicians.
• Occlusion Limitations: Dense terminal hair occlusions across the lesion boundary may distort the perilesional ring contrast calculation. Pre-processing with digital hair removal (e.g. DullRazor) is strongly recommended.
• Amelanotic Lesions: Highly amelanotic non-pigmented lesions on Fitzpatrick Type I skin exhibit negligible optical contrast ($\Delta E^*_{ab} < 8.0$), where multi-modal or clinical dermoscopy priors are advised.

---

Quickstart & Usage

1. Installation

git clone https://github.com/raju-sah/MedSAM-For-Skin-Segmentation.git
cd MedSAM-For-Skin-Segmentation
pip install -r requirements.txt

2. Standalone CLI Inference

# Single image with auto-prompt detection
python inference.py --image path/to/sample.jpg --output_dir output/ --model cg_adapter

# Custom clinician bounding box (X1 Y1 X2 Y2)
python inference.py --image path/to/sample.jpg --bbox 120 80 450 380 --output_dir output/

# Batch process entire folder
python inference.py --input_dir demo_samples/ --output_dir output/ --model cg_adapter

3. Interactive Web Demo

# Launch modern browser demonstration UI
uvicorn web_demo.app:app --host 127.0.0.1 --port 7860
# Navigate to http://127.0.0.1:7860 in your browser

4. Programmatic PyTorch API

import cv2
import torch
from src.eval.inference_utils import load_model, run_inference

# Initialize model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, _ = load_model(model_name="cg_adapter", device=device)

# Load RGB image and specify bounding box
image_bgr = cv2.imread("demo_samples/sample_dark_fst_v.jpg")
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
bbox = (50, 114, 109, 155)

# Forward inference
result = run_inference(model, image_rgb, bbox, device=device)
binary_mask = result["mask"]
telemetry = result["contrast_info"]

print(f"DeltaE*ab: {telemetry['delta_e']} | Skin Tone: {telemetry['estimated_fst']} | Gamma: {telemetry['gamma_factor']}")

---

Checksum Verification

a09fcf2808dc101c87ea255b79a9830b4cb4c1e2e7ad1bd36e9c6756cbf73b38  best_cg_adapter_model.pth
8e8e4a04acc7aa05236b34ffceb1f1168d7d48e86b15c232117e41bcca28ead3  best_standard_adapter_model.pth
e8c107d3c34e8122931909adcbb05410be42c1d79665fa61a735c6c7c82ea2ac  best_lora_model.pth
695430f33f9772da50972ed3a363671f0625a5081804336264be2d02f4ad896f  best_decoder_only_model.pth
34b34b78c1d18cb8c6bf84cf9c00e135d6d6c965699f3c0e31ef1bc9dcb5be74  medsam_vit_b.pth

---

Citation

@inproceedings{cg_medsam_2026,
  title     = {Contrast-Gated Parameter-Efficient Adaptation of MedSAM for Skin-Tone-Robust Lesion Segmentation},
  author    = {Anonymous},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  year      = {2026},
  note      = {Paper ID: 1042}
}