Manchester Metropolitan University•Jan 2026 – Aug 2026•vision

Cross-Domain Melanoma Detection via Unsupervised Domain Adaptation

MSc Dissertation research at Manchester Metropolitan University (Supervised by Dr. Taysir Dyhoum) establishing the first unified benchmark of 9 unsupervised domain adaptation (UDA) methods across adversarial, statistical, and distillation families on an EfficientNet-B2 backbone to overcome domain shift between dermoscopy and clinical photography for melanoma detection.

Executive Overview & Key Engineering Deliverables

Clinical Imaging Disparity: Addressed the diagnostic domain gap between controlled dermoscopic imaging and ordinary clinical photographs, enabling dermatological AI generalisation to non-specialist and smartphone settings.
9 UDA Architectures Evaluated: Benchmarked DANN, CDAN, MCD, FFTAT-inspired (adversarial), Deep CORAL, MMD, JAN (statistical), and EKDA-inspired, CRPL-inspired (distillation/prototypes) under a unified evaluation framework.
Extensive 5-Seed Empirical Protocol: Evaluated across 3 cross-domain scenarios (Derm7pt, ISIC 2018 to PAD-UFES-20 with 34:1 imbalance, Milk10k) on 6 held-out test sets across AUROC, AUPRC, Melanoma Recall, and Expected Calibration Error (ECE).
Sensitivity vs. Calibration Discovery: Proved that high-recall methods (MCD recall 66.7%, CRPL recall 63.4%) incur severe calibration penalties (ECE > 24%), whereas EKDA distillation gains recall while achieving best-in-class calibration (ECE 10.88%).

PyTorchEfficientNet-B2Unsupervised Domain AdaptationGrad-CAMMedical Image AnalysisDeep CORALDANNEKDA DistillationCUDAPython

UDA Methods

9 Algorithms

Backbone

EfficientNet-B2

Top AUROC

71.8% (CRPL)

Top ECE

10.9% (EKDA)

LIVE SYSTEM SIMULATION & INTERACTIVE PREVIEW

STATUS: ACTIVE WEBGL/SIM

UNSUPERVISED DOMAIN ADAPTATION // MELANOMA DETECTION STUDIO

EFFICIENTNET-B2 • 9 UDA BENCHMARKS • GRAD-CAM EXPLAINABILITY (MMU 2026)

UNSUPERVISED DOMAIN ADAPTATION // MELANOMA STUDIOMMU DISSERTATION

EFFICIENTNET-B2 BACKBONE • 9 UDA BENCHMARKS • MMU RESEARCH

SCENARIO:

SAME-PATIENT PAIREDPure modality shift. Identical patient lesions photographed with polarized dermoscopy and clinical camera.

UDA METHOD SELECTOR (9 BENCHMARKED ARCHITECTURES):Backbone: EfficientNet-B2 (1,408-d features)

DISCRIMINATION (AUROC)

71.75%

95% CI: [69.8, 73.5]

MINORITY PRECISION (AUPRC)

31.21%

95% CI: [30.2, 34.6]

MELANOMA RECALL (0.5 THRESH)

63.40%

95% CI: [60.5, 66.2]

CALIBRATION ERROR (ECE)

24.86%

Lower is better (Reliability)

INFERENCE LATENCY

14.8ms

Single Image Forward Pass

Pause Loop

Reset Epoch

EPOCH: 75%

CONVERGED // DOMAIN INVARIANCE ACHIEVED

DUAL-STREAM UNIFIED ADAPTATION PIPELINE // EFFICIENTNET-B2 BACKBONE

Loss Schedule λ(p): 0.999•H-Divergence: 0.54

Mathematical Objective: L_total = L_bce + λ_proto · Σ_c ||z_target - μ_c||²

Domain Gap Reduction: 64%

ACTIVE UDA MECHANISM

Prototype clustering loss pulling target features toward class-conditional source centroid anchors.

distillation family

CLINICAL MOTIVATION

Deep neural networks trained on high-resolution dermoscopy experience severe generalization collapse on smartphone photography due to sensor noise, uneven ambient lighting, and patient skin tone variation.

EMPIRICAL BENCHMARK FINDING

CRPL-Inspired Prototypes achieves 71.8% AUROC and 63.4% recall on held-out target data without requiring any target training labels.

FULL-STACK SYSTEM ARCHITECTURE SPECIFICATION

CLIENT & UI TIER

Interactive diagnostic visualizer with multi-scenario comparison across 6 held-out test sets (D7D, D7C, M10D, M10C, ISIC18, PAD).
Layer-wise Grad-CAM activation heatmap visualizer overlaying lesion boundaries and diagnosing clinical artifact vulnerabilities.
Reliability diagrams and calibration error plots evaluating Expected Calibration Error (ECE) across decision thresholds.
2D latent feature space scatter visualizer projecting source vs target distribution alignment in full vs ablated modes.

API & REAL-TIME ENGINE

Shared EfficientNet-B2 feature extractor (G_theta) initialized with ImageNet pretraining, extracting 1408-dimensional feature representations.
Dual-loss objective function (L_total = L_cls + lambda * L_ada) with binary cross-entropy on source labels and unsupervised domain alignment on target samples.
Adversarial domain discriminators with Gradient Reversal Layers (GRL) and multilinear conditioning (CDAN).
Statistical second-order covariance alignment (Deep CORAL) and multi-kernel Maximum Mean Discrepancy (MMD / JAN).
Teacher-student exponential moving average knowledge distillation (EKDA) and cluster-preserving prototypes (CRPL).

DATA & INFRASTRUCTURE

ISIC 2018 benchmark dermoscopic archive (10,015 images collapsed into binary melanoma vs other).
PAD-UFES-20 clinical photography collection (2,298 smartphone photos with severe 34:1 class imbalance).
Derm7pt paired dataset (1,011 dermoscopic and clinical lesion pairs across identical patients).
Milk10k paired dataset (5,240 paired dermoscopic and clinical images).
PyTorch 2.x analytical runtime with CUDA acceleration and mixed-precision forward passes (~14.2ms GPU latency).
2,000-iteration stratified image-clustered bootstrap uncertainty estimation pipeline.
Standardized 5-seed reproducibility harness with deterministic seeds and fixed hardware configurations.

ROLE-BASED WORKFLOW SPECIFICATION (RBAC)

Lead Deep Learning & Computer Vision ResearcherRESEARCH & ALGORITHM DESIGN

▸Formulated the benchmarking research design evaluating 9 UDA algorithms on EfficientNet-B2 under unified training protocols
▸Engineered custom PyTorch modules for GRL domain confusion, covariance alignment, MMD kernels, and distillation loss functions
▸Designed the binary melanoma classification pipeline with class-balanced weighting and stratified data splitting
▸Conducted 2,000-iteration stratified image-clustered bootstrap uncertainty estimation

Clinical AI & Diagnostic Evaluation SpecialistCLINICAL EVALUATION & METRICS

▸Evaluated cross-modality diagnostic shifts between dermoscopy and clinical smartphone photography
▸Analyzed the critical trade-off between melanoma sensitivity (recall) and clinical overconfidence (ECE)
▸Investigated severe class imbalance challenges using PAD-UFES-20 (34:1 ratio) via minority-sensitive AUPRC
▸Validated model generalisation across 6 held-out test datasets without target labels

Model Explainability & Grad-CAM AuditorEXPLAINABILITY & RELIABILITY

▸Implemented Grad-CAM activation mapping to audit neural attention across lesion boundaries
▸Identified spurious feature artifacts including hair follicles, skin folds, and clinical lighting variations
▸Audited five-seed pooled confusion matrices to identify false negatives and low-confidence melanoma cases
▸Conducted inference latency profiling (~14.2ms forward pass on GPU) for prospective point-of-care feasibility

Key Technical Achievements

01First Unified UDA Cross-Family Benchmark: Compared adversarial (DANN, CDAN, MCD, FFTAT), statistical (Deep CORAL, MMD, JAN), and distillation (EKDA, CRPL) methods under a single identical evaluation protocol.
02Discovery of Sensitivity vs. Calibration Trade-off: Revealed that methods pushing target decision confidence (MCD recall 66.74%, CRPL recall 63.40%) incur the highest calibration penalties (ECE 30.55% and 24.86%), while EKDA distillation preserves calibration (ECE 10.88%).
03Severe Imbalance Robustness: Evaluated real-world smartphone photos from PAD-UFES-20 under 34:1 class imbalance, demonstrating Deep CORAL leads minority-class AUPRC (33.93%).
04Qualitative Grad-CAM Feature Auditing: Visualized neural activation maps across lesion borders, discovering that successful adaptation refocuses attention from spurious lighting and skin wrinkles to pigmented lesion structures.

Measurable Impact & Production Verification

Successfully completed and defended MSc Dissertation in Data Science at Manchester Metropolitan University. Proved that unsupervised domain adaptation significantly mitigates cross-modality domain shift between dermoscopy and clinical photography, revealing crucial trade-offs between melanoma recall and calibration error, and providing the first unified benchmark of adversarial, statistical, and distillation UDA methods for dermatological AI.

Key Production Deliverables & Verification

Unified UDA Benchmark Protocol

9 adaptation algorithms evaluated across 3 cross-modality scenarios with 5 random seeds on 6 held-out test sets.

Multi-Metric Diagnostic Evaluation

Comprehensive assessment using AUROC, minority-class AUPRC, threshold-bound Melanoma Recall, and Expected Calibration Error (ECE).

Grad-CAM Explainability & Attention Auditing

Qualitative attention heatmaps auditing lesion boundary localization and diagnosing clinical artifact vulnerabilities.

Stratified Bootstrap Uncertainty

2,000 image-clustered bootstrap iterations establishing robust empirical confidence intervals for all macro-averages.