JOURNAL OF INFORMATION SYSTEM RESEARCH (JOSH)
Vol 8 No 1 (2026): October 2026

True Structured Network Surgery of EfficientNet-B3 with Focal Loss for Highly Imbalanced Dermoscopic Image Analysis

Argha Orion Silitonga (Universitas Klabat, Minahasa Utara)



Article Info

Publish Date
29 Oct 2026

Abstract

Skin tumor classification relies heavily on deep convolutional neural networks, which are computationally expensive for edge deployment. This study proposes a resource-efficient pipeline using true structured network surgery on an EfficientNet-B3 architecture. To address the extreme class imbalance in the ISIC dataset, we implemented minority class oversampling alongside a custom Binary Focal Loss function. Our experiments evaluated 0%, 20%, and 50% structural channel pruning utilizing L2-norm magnitude importance. The baseline model achieved an Area Under the Curve (AUC) of 0.9314 with 10.70 million parameters. At 50% pruning, the parameter footprint dropped by 74% to 2.79 million, and floating-point operations (FLOPs) decreased by 71% to 0.55 G. Despite these reductions, latency metrics indicated a hardware execution plateau on an NVIDIA RTX 3060 Laptop GPU, stabilizing at approximately 12.93 ms per image. By dynamically optimizing the decision threshold using Youden's J statistic, the 50% pruned model maintained a high malignant sensitivity (recall) of 81.59%, despite an overall accuracy reduction to 76.69%. These results demonstrate that combining physical tensor pruning with focal loss yields highly compressed models capable of robust malignant lesion recall. Furthermore, post-hoc visual interpretability via Grad-CAM verified that these compressed architectures retained their clinical diagnostic integrity by accurately localizing genuine pathological features. The primary contributions of this research include the demonstration of true structural channel pruning to reduce model parameters by 74% for edge deployment, the preservation of an 81.59% malignant sensitivity in a highly compressed network using Binary Focal Loss and dynamic thresholding via Youden's J statistic, and the verification of post-pruning diagnostic integrity using Grad-CAM. This pipeline presents a viable pathway for deploying lightweight, accurate dermatological diagnostic tools in resource-constrained clinical environments.

Copyrights © 2026






Journal Info

Abbrev

josh

Publisher

Subject

Computer Science & IT Decision Sciences, Operations Research & Management

Description

Artikel yang dimuat melalui proses Blind Review oleh Jurnal JOSH, dengan mempertimbangkan antara lain: terpenuhinya persyaratan baku publikasi jurnal, metodologi riset yang digunakan, dan signifikansi kontribusi hasil riset terhadap pengembangan keilmuan bidang teknologi dan informasi. Fokus Journal ...