← Back to Publications List

ADFNet: Adaptive Dual-Attention Fusion in EfficientNet-B4 for Robust Multi-Class Dermoscopic Skin Lesion Classification

Students & Supervisors

Student Authors
Tanvir Mahtab Taneem
Bachelor of Science in Computer Science & Engineering, FST
Supervisors
Tanvir Ahmed
Assistant Professor, Faculty, FST
Rifath Mahmud
Assistant Professor, Faculty, FST
Syeda Anika Tasnim
Assistant Professor, Faculty, FST

Abstract

Dermoscopic skin lesion classification demands mod els that handle severe class imbalance and supply inter pretable predictions. Existing attention-based approaches follow the CBAM convention of fixed channel-before-spatial ordering, yet no prior work has validated this ordering on dermoscopic data. We present ADFNet, built around an Adaptive Dual Attention Module (ADAM) that replaces the fixed ordering with a single learnable scalar α, optimised end-to-end alongside all other parameters. When trained on HAM10000 with Focal Loss, MixUp, and a class-balanced sampler, α converges to 0.4520, indicating that spatial localisation marginally outweighs channel recalibration on this domain. Under 8-variant test-time augmentation, ADFNet achieves 79.24% accuracy and 0.9617 macro AUC on the held-out test; a soft-vote ensemble with the CBAM-free baseline yields 81.14% accuracy and 0.9705 macro AUC. After fine-tuning on ISIC 2019, the ensemble reaches 66.30% across five overlapping classes. Grad-CAM maps confirm lesion-centred attention in both datasets. The interpretable α is a clinically readable signal that no fixed-order architecture can produce.

Keywords

Index Terms—Skin lesion classification dermoscopy adaptive attention EfficientNet-B4 HAM10000 ISIC 2019 Grad-CAM class imbalance interpretable AI.

Publication Details

  • Type of Publication:
  • Conference Name: IEEE SPICSCON 2026 IEEE International Conference on Signal Processing, Information, Communication and Systems 2026
  • Date of Conference: 13/08/2026 - 13/08/2026
  • Venue: Bangladesh Army University of Engineering & Technology (BAUET), Qadirabad, Natore-6431, Bangladesh.
  • Organizer: IEEE Bangladesh Section (IEEE BDS) and the IEEE Signal Processing Society (SPS) Bangladesh Chapter.
View on Publisher