DermaSense

CNN-based skin lesion classification for early skin cancer detection — a research-grade diagnostic-support prototype combining a 4-model CNN ensemble, post-hoc probability calibration, Grad-CAM saliency explainability, and a clinical-style Flutter mobile/web interface.

Status: v2.1 — UI Redesign · see Releases License: Code MIT · Model weights CC BY-NC 4.0 (see LICENSE, LICENSE-MODELS.md) Not a clinical device. Research and educational use only.

Overview

DermaSense classifies dermatoscopic images into seven HAM10000 skin lesion categories and presents calibrated confidence values, a Top-3 differential, and on-demand Grad-CAM attention overlays through a Flutter interface. The system uses four ImageNet-pretrained backbones (ResNet50, DenseNet121, EfficientNet-B0, MobileNetV3 Small), a temperature-scaled calibration layer, a weighted ensemble with explicit disagreement detection, and a FastAPI backend exposing five HTTP endpoints.

Highlights

• 4-model weighted ensemble with cross-model disagreement detection surfaced as a first-class UI signal (indeterminate risk state + banner).
• Post-hoc temperature scaling per model, fit on the HAM10000 validation split and verified on the test split; deployed ResNet50 v2 reaches post-calibration test ECE 0.0248.
• Grad-CAM saliency for single-model and per-model ensemble views; lazy fetch with overlay caching in the app.
• ResNet50 v2 retraining (focal loss + class-balanced sampler) raises melanoma recall from 54.8% to 73.40% on the HAM10000 test split.
• External validation on a 4,353-image HAM-disjoint subset of ISIC 2019, with the resulting in-distribution / out-of-distribution gap surfaced as a Known Limitations panel inside the app.
• Clinical-style Flutter UI built on a centralised design-token system (institutional navy + teal, risk-state accents, hairline borders, no shadows).

Performance summary (HAM10000 test split)

Model	Test Acc	Macro F1	mel recall
MobileNetV3 Small	67.76%	57.26%	—
EfficientNet-B0	77.45%	64.77%	—
DenseNet121	79.64%	68.96%	—
ResNet50 (v1, baseline)	80.22%	69.03%	54.8%
ResNet50 v2 (focal + sampler, deployed single)	—	70.08%	73.40%
4-model weighted ensemble	—	74.10%	—

External validation on ISIC 2019 (HAM-disjoint subset, 4,353 images): single-model mel recall drops 73.40% → 37.09%; ensemble macro F1 drops 74.10% → 41.24%. This gap is documented and surfaced to users.

Quick start

See REPRODUCE.md for the full end-to-end reproduction recipe (clone, download released model weights via gh release download v2.1, run backend + Flutter Web). Minimum requirements: Python 3.10+, Flutter 3.16+, gh CLI authenticated.

Architecture

• Backend (src/skinlesion/, FastAPI on port 8126):
  • GET /health — service status
  • POST /predict — single-model prediction (ResNet50 v2 + calibration)
  • POST /predict-ensemble — 4-model weighted ensemble + disagreement detection
  • POST /cam — Grad-CAM heatmap for the deployed single model
  • Per-model CAM returned as part of the ensemble payload
• Mobile app (mobile_app/, Flutter — runs as web in Chrome for demonstration):
  • Five screens: Home, Analysis Setup, Analysis Result (single / ensemble), Model Performance, About & Safety
  • Centralised design tokens (mobile_app/lib/theme/design_tokens.dart)
  • Risk-state-coloured result hero · 200 ms cross-fade Grad-CAM toggle · 2×2 per-model ensemble grid with shared lightbox

Repository layout

src/skinlesion/         # Python: training, inference, calibration, Grad-CAM
mobile_app/             # Flutter app (web/mobile)
configs/                # Training configs (v1 baselines + v2 focal+sampler)
runs/                   # Per-model checkpoints + calibration JSONs
runs_v2/                # Phase C v2 training experiment outputs
docs/                   # cam_design.md, calibration_report.md, ui_redesign/, figures/
scripts/                # Smoke tests + ad-hoc utilities
REPRODUCE.md            # End-to-end reproduction recipe
LICENSE                 # MIT (source code)
LICENSE-MODELS.md       # CC BY-NC 4.0 (released model weights)
CITATION.cff            # Citation metadata

Known limitations

The deployed application surfaces these in a dedicated Known Limitations panel; the report covers them in detail. Briefly:

• External-validation drop. v2 single-model mel recall falls 73.40% → 37.09% on ISIC 2019 HAM-disjoint; ensemble macro F1 falls 74.10% → 41.24%. In-distribution metrics do not transfer.
• Calibration scope. Temperature scaling is fit and verified on HAM10000 only; not characterised for out-of-distribution images.
• Class imbalance. Rare classes (df, vasc) retain lower-confidence boundaries even after v2.
• Imaging conditions. All training data are dermatoscopic; phone-camera photos without dermatoscopic contact are out of distribution.
• Not a clinical device. No regulatory validation. Predictions must not be acted on without qualified clinician review.

License

• Source code: MIT
• Released model weights (the twelve *.pt and *calibration.json assets attached to GitHub Releases): CC BY-NC 4.0, inherited from HAM10000 and ISIC 2019.

Citation

If you use this project or its released model weights, please cite via the GitHub "Cite this repository" button (powered by CITATION.cff) or:

Liu, J. DermaSense — CNN-Based Skin Lesion Classification for Early Skin Cancer Detection. MSc coursework project (UTS 42028), 2026. Release v2.1. https://github.com/Cureeeeeeee/CNN-Based-Skin-Lesion-Classification-for-Early-Skin-Cancer-Detection

Datasets

• HAM10000: Tschandl, Rosendahl, Kittler. "The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions." Scientific Data, vol. 5, 180161, 2018. Licensed CC BY-NC 4.0.
• ISIC 2019: Combalia et al. "BCN20000: Dermoscopic Lesions in the Wild." arXiv:1908.02288, 2019. Licensed CC BY-NC 4.0.

Acknowledgement

Built as the UTS 42028 (Deep Learning and Convolutional Neural Network) Assignment-3 final project. Iterative development assisted by Anthropic's Claude (commits include Co-Authored-By: Claude trailer for transparency).