Publications
Manuscripts, research outputs, and ongoing scholarly work.
Manuscripts Under Review
Generalizable Ensemble Deep Learning for Dermoscopic Skin-Lesion Classification: Internal Evaluation on HAM10000 and External Evaluation on ISIC 2019
Status: Under review (2026)
Author: Md Naim Hassan Saykat
This study investigates the generalizability of deep learning models for multiclass dermoscopic skin-lesion classification using harmonized internal evaluation on HAM10000 and external evaluation on ISIC 2019.
Seven architectures are evaluated: a conventional CNN, ResNet-50, DenseNet-121, EfficientNet-B3, ConvNeXt-Tiny, MobileNetV3-Large, and ViT-B/16. Their class-probability outputs are additionally combined using an equal-weight soft-voting ensemble.
The evaluation includes classification and discrimination metrics, calibration analysis, bootstrap confidence intervals, paired statistical comparisons, and Grad-CAM-based qualitative interpretability analysis. The study emphasizes external generalization and the limitations of relying exclusively on internal performance.
- Seven deep learning architectures and probability-level ensemble learning
- HAM10000 internal evaluation and ISIC 2019 external evaluation
- Model calibration and bootstrap uncertainty analysis
- Paired statistical model comparisons
- Grad-CAM-based interpretability analysis
- Reproducible medical-image classification workflow
Research Outputs
Code, evaluation workflows, curated result tables, figures, and supporting reproducibility materials for the skin-lesion classification study are publicly available through the associated GitHub repository and persistent Zenodo archive.
In Preparation
Edge-Ready Skin Lesion Classification with Efficient Deep Ensembles and Latency-Aware Design: A Web-Based Decision-Support Prototype
Status: Research in preparation
Planned submission: 2026-27 academic year
This planned study extends the skin-lesion classification research toward computationally efficient deployment. It will investigate the trade-offs among predictive performance, model size, inference latency, and computational requirements for resource-constrained environments.
The planned methodology includes efficient model and ensemble configurations, model compression, quantization, inference benchmarking, and Grad-CAM-based interpretability. Evaluation will focus on the performance-efficiency trade-off rather than predictive performance alone.
A lightweight web-based research prototype is also planned to demonstrate model inference and visual explanation workflows. The prototype is intended for research and demonstration purposes and is not a clinical diagnostic system.
Research Profile
For persistent researcher identification and an up-to-date record of scholarly outputs, see my ORCID profile.
Publication status and bibliographic information will be updated as manuscripts progress through peer review, revision, publication, and archival release.