Melanoma Skin Lesion Classifier

A fine-tuned ResNet-18 classifier trained on the HAM10000 dermoscopy dataset. Given a skin lesion image, the model predicts Melanoma vs. Non-Melanoma and produces a visual analysis report with confidence percentages, a Grad-CAM heatmap, and a lesion bounding box.

---

Project Structure

Melanoma-Skin-Classifier/
├── Training_Model.py       # Full training pipeline
├── Inference.py            # Single-image analysis with visual report
├── Augmentation.py         # Transform definitions (train + val)
├── Load_Image.py           # Dataset class with subfolder support
├── Load_PreTrained.py      # Model loading utility
├── requirements.txt        # Python dependencies
├── checkpoints/            # Saved during training (not tracked by git)
│   ├── resnet18_best.pth       # Best checkpoint (lowest val loss)
│   ├── resnet18_final.pth      # Final epoch checkpoint
│   └── training_history.json  # Per-epoch metrics
├── results/                # Inference reports (not tracked by git)
│   └── <image>_analysis.png
└── data/                   # Dataset (not tracked by git)
    └── images/
        ├── HAM10000_metadata.csv
        ├── HAM10000_images_part_1/
        └── HAM10000_images_part_2/

---

Setup

pip install -r requirements.txt

---

Training

Features

Feature	Detail
Base model	ResNet-18 (ImageNet pretrained)
Classes	Melanoma (mel) · Non-Melanoma (all other dx)
Dataset split	80 % train / 20 % validation (stratified)
Class imbalance	WeightedRandomSampler (inverse class frequency)
Augmentation	Flip, rotation ±20°, colour jitter
Optimiser	Adam — lr 1e-4
Loss	CrossEntropyLoss
Early stopping	Patience = 3 epochs (monitors val loss)
Metrics	Accuracy, AUC-ROC, Precision, Recall, F1
GPU support	Auto-detects CUDA, falls back to CPU

Run Training

Default run (expects data at data/images/):

python Training_Model.py

Custom paths:

python Training_Model.py \
  --img-dir  /path/to/HAM10000/images \
  --csv      /path/to/HAM10000_metadata.csv \
  --ckpt-dir /path/to/save/checkpoints \
  --epochs   20 \
  --patience 5

All arguments:

--img-dir     Root directory of HAM10000 images   (default: data/images)
--csv         Path to HAM10000_metadata.csv        (default: data/images/HAM10000_metadata.csv)
--ckpt-dir    Directory to save checkpoints        (default: checkpoints)
--epochs      Max training epochs                  (default: 10)
--patience    Early-stopping patience              (default: 3)
--batch-size  Batch size                           (default: 32)
--lr          Adam learning rate                   (default: 1e-4)

---

Inference — Submitting a Skin Lesion Image

python Inference.py --image path/to/lesion.jpg

Optional arguments:

--checkpoint   Path to .pth model file   (default: checkpoints/resnet18_best.pth)
--output       Path to save report PNG   (default: results/<name>_analysis.png)

What the Report Shows

The output is a 4-panel PNG report:

Panel	Description
Original Image	Input lesion resized to 224×224
Grad-CAM Heatmap	Jet-coloured overlay — red/yellow regions most influenced the prediction
Lesion Detection	Original image with a bounding box drawn around the highest-activation region, labelled with prediction + %
Diagnosis	Prediction verdict · Melanoma % bar · Non-Melanoma % bar · Confidence score

Example Console Output

══════════════════════════════════════════════
  Skin Lesion Analysis Report
══════════════════════════════════════════════
  Image        : ISIC_0024306.jpg
  Prediction   : Melanoma
  Melanoma     : 87.3%
  Non-Melanoma : 12.7%
  Confidence   : 87.3%  (High Confidence)
  Bounding Box : x1=48 y1=52 x2=178 y2=191
  Report saved : results/ISIC_0024306_analysis.png
══════════════════════════════════════════════

⚕  Results are for medical reference only.

Bounding Box

The bounding box is derived from the Grad-CAM activation map: the heatmap is thresholded at 45 %, the largest contour is found with OpenCV, and cv2.boundingRect returns [x1, y1, x2, y2] coordinates on the 224×224 image space.

---

Dataset

HAM10000 — Human Against Machine with 10000 training images
Tschandl, P. et al. (2018). The HAM10000 dataset. Scientific Data.
Download: https://www.kaggle.com/datasets/kmader/skin-lesion-analysis-toward-melanoma-detection

---

⚕ This tool is intended for research and educational purposes only and does not constitute medical advice.