Attention-Enhanced DenseNet201 for Nine-Class Skin Lesion Classification: A Comparative Study with Deployment Considerations

Melanoma accounts for over 25,000 annual deaths, with 5-year survival dropping to 20% when diagnosis is delayed. Although expert dermatologists achieve 90–95% diagnostic accuracy, access remains limited in resource-constrained settings, necessitating reliable automated screening support. Existing deep learning approaches frequently optimize overall accuracy while neglecting melanoma-specific recall, producing clinically dangerous false-negative rates. We present an attention-enhanced DenseNet201 architecture incorporating a Convolutional Block Attention Module (CBAM) with selective layer unfreezing, trained on a combined nine-class dermoscopic corpus built from four public International Skin Imaging Collaboration (ISIC)-family releases (14,400 training/3600 validation/4500 test images) at a reduced 100 × 75 input resolution. The preferred configuration (Model B: DenseNet201 with a single image-conditioned multi-head attention layer) achieved 93.72% validation accuracy, 80.65% melanoma recall, 90.03% melanoma precision, and a weighted F1 of 0.9372. Fivefold stratified cross-validation confirmed the robustness of its performance (93.72% ± 0.37% accuracy; 83.50% ± 0.53% melanoma recall on a held-out stratified subsample). Under an identical training protocol, the proposed model outperformed a non-attention DenseNet201 baseline, with McNemar’s test confirming statistical significance (χ2=342.7, p<0.001). ResNet50V2 and EfficientNetB0 were originally trained under a non-matched, single-fixed-learning-rate protocol; a controlled re-evaluation of ResNet50V2 with a proper backbone-specific learning-rate search (selected rate 0.001) showed that its validation accuracy improved from 56.58% to 88.17%, confirming that the original protocol mismatch, not the architecture itself, was responsible for the earlier collapse. A parallel controlled re-evaluation of EfficientNetB0 was initiated but interrupted before completion and is reported as still pending. Ablation studies established that selective layer unfreezing (93.7%) outperforms frozen transfer learning (84.6%) and full fine-tuning (91.5%). The model comprises 19.65 M parameters and has 105 ± 12 ms inference latency on an NVIDIA (Santa Clara, CA, USA) Tesla T4 (9.5 frames per second, FPS). We treat the 100 × 75 resolution as an unresolved, consequential limitation rather than a validated design choice, restrict our comparative claims accordingly, and position the system as a screening-support aid rather than a standalone diagnostic device.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-16
DOI
https://doi.org/10.3390/electronics15225329
Primary Topic
Cutaneous Melanoma Detection and Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Attention-Enhanced DenseNet201 for Nine-Class Skin Lesion Classification: A Comparative Study with Deployment Considerations

Hina Syed, Hassan A. Ahmed, Syed Hamza Ahmed, Brigid Campbell
3 links
Electronics
Cutaneous Melanoma Detection and Management
article

Attention-Enhanced DenseNet201 for Nine-Class Skin Lesion Classification: A Comparative Study with Deployment Considerations

Hina Syed, Hassan A. Ahmed, Syed Hamza Ahmed, Brigid Campbell
article en

Abstract

Melanoma accounts for over 25,000 annual deaths, with 5-year survival dropping to 20% when diagnosis is delayed. Although expert dermatologists achieve 90–95% diagnostic accuracy, access remains limited in resource-constrained settings, necessitating reliable automated screening support. Existing deep learning approaches frequently optimize overall accuracy while neglecting melanoma-specific recall, producing clinically dangerous false-negative rates. We present an attention-enhanced DenseNet201 architecture incorporating a Convolutional Block Attention Module (CBAM) with selective layer unfreezing, trained on a combined nine-class dermoscopic corpus built from four public International Skin Imaging Collaboration (ISIC)-family releases (14,400 training/3600 validation/4500 test images) at a reduced 100 × 75 input resolution. The preferred configuration (Model B: DenseNet201 with a single image-conditioned multi-head attention layer) achieved 93.72% validation accuracy, 80.65% melanoma recall, 90.03% melanoma precision, and a weighted F1 of 0.9372. Fivefold stratified cross-validation confirmed the robustness of its performance (93.72% ± 0.37% accuracy; 83.50% ± 0.53% melanoma recall on a held-out stratified subsample). Under an identical training protocol, the proposed model outperformed a non-attention DenseNet201 baseline, with McNemar’s test confirming statistical significance (χ2=342.7, p<0.001). ResNet50V2 and EfficientNetB0 were originally trained under a non-matched, single-fixed-learning-rate protocol; a controlled re-evaluation of ResNet50V2 with a proper backbone-specific learning-rate search (selected rate 0.001) showed that its validation accuracy improved from 56.58% to 88.17%, confirming that the original protocol mismatch, not the architecture itself, was responsible for the earlier collapse. A parallel controlled re-evaluation of EfficientNetB0 was initiated but interrupted before completion and is reported as still pending. Ablation studies established that selective layer unfreezing (93.7%) outperforms frozen transfer learning (84.6%) and full fine-tuning (91.5%). The model comprises 19.65 M parameters and has 105 ± 12 ms inference latency on an NVIDIA (Santa Clara, CA, USA) Tesla T4 (9.5 frames per second, FPS). We treat the 100 × 75 resolution as an unresolved, consequential limitation rather than a validated design choice, restrict our comparative claims accordingly, and position the system as a screening-support aid rather than a standalone diagnostic device.

ElectronicsVol. 15(22)
Cleveland State University (US), Riphah International University (PK)
Good health and well-being
Cutaneous Melanoma Detection and Management
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