Skin Lesion Classification using CNN

Skin cancer is the most common cancer worldwide, and dermatoscopy — magnified imaging of pigmented lesions — is how dermatologists triage them. Deep learning reached dermatologist-level classification on the HAM10000 benchmark (Tschandl et al., 10,015 dermatoscopic images, 7 diagnostic categories), making it one of the most cited medical-imaging results. This project reproduces that pipeline: an ImageNet-pretrained EfficientNet-B0 fine-tuned on HAM10000 with class weights for the heavy imbalance toward benign nevi. A live web demo makes the dermatology tangible — upload a close-up lesion photo and the app segments it by Otsu thresholding, then computes real ABCD features from the pixels: asymmetry by mirror-overlap, border irregularity by compactness, color by distinct cluster count, diameter by lesion extent. Because the demo states plainly that the ABCD score is a transparent rule and the project is an educational prototype, the student can defend the scope honestly in the viva.

1. HAM10000 images are loaded with their 7-class labels, stratified-split, normalized and augmented with documented transforms (rotation, flip, color jitter).
2. EfficientNet-B0 (ImageNet weights) is fine-tuned with a 256-unit dense head, dropout 0.4 and class-weighted cross-entropy to counter the nv majority.
3. The held-out split is evaluated once: balanced accuracy, per-class precision/recall/F1, top-2 accuracy and the confusion matrix are generated for the report.
4. In the web demo, an uploaded photo is segmented by Otsu thresholding on the brown-pigment channel to isolate the lesion.
5. Real ABCD features are computed from the mask: asymmetry by mirror non-overlap, border by 8 − compactness×8, color by distinct quantized clusters, diameter by lesion width fraction.
6. The transparent ABCD rule combines them into a triage score with clinical bands; the built-to-order CNN replaces the rule with learned 7-class probabilities.

Tech stack:

• Python 3, TensorFlow/Keras (EfficientNet)
• NumPy, OpenCV, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML5 canvas + JavaScript (live ABCD demo)
• HAM10000 dataset

Dataset & model details

• Dataset: HAM10000 (Tschandl et al., Harvard / Medical University of Vienna) — 10,015 dermatoscopic images across 7 categories: akiec, bcc, bkl, df, mel, nv, vasc. Ground truth from histopathology and expert consensus; the nv (benign nevus) class dominates and is handled with class weights.
• Task: Lesion classification; input = 224×224×3 dermatoscopic image, output = probability distribution over the 7 HAM10000 categories.
• Model: EfficientNet-B0 (ImageNet initialization) → global average pooling → Dense(256, ReLU) → Dropout(0.4) → Dense(7, softmax); class-weighted cross-entropy, fine-tuning schedule documented.
• Metrics: Balanced accuracy ≥ 80% (design target), melanoma recall ≥ 85% (design target), top-2 accuracy reported (evaluated), per-class precision/recall/F1. No metric is claimed as measured until the training run executes for the order.

• [7-class CNN classifier] EfficientNet-B0 with ImageNet initialization, global pooling, a 256-unit dense head and 7-way softmax over the HAM10000 categories — architecture documented layer by layer.
• [Live ABCD analyzer demo] Upload a lesion close-up; the app segments it by Otsu thresholding on the brown-pigment channel and computes real asymmetry, border, color and diameter from the actual pixels.
• [Transparent ABCD scoring] The clinical scoring rule (score = 1.3·A + 0.1·B + 0.5·C + 0.5·D) applied to the measured features with the published interpretation bands — clearly labeled as the rule, not the CNN.
• [ABCD Lab] Each letter of the dermatology rule explained with the exact computation used, so the viva can walk through asymmetry, border, color and diameter precisely.
• [Full training notebook] HAM10000 loading, augmentation, class-weighted loss, fine-tuning schedule and evaluation in one reproducible Jupyter notebook.
• [Per-class metrics] Precision, recall and F1 for each of the seven categories plus the confusion matrix, so the viva can discuss exactly which lesions confuse the model.
• [Imbalance handling] The nv-majority imbalance addressed with class weights and documented augmentation, with the effect measured and reported.

• Complete training & evaluation Jupyter notebook on HAM10000
• Fine-tuned EfficientNet model file with preprocessing code
• Live ABCD lesion analyzer + ABCD Lab + System Report web demo
• Confusion matrix, per-class metrics and training-curve plots
• Project report PDF (dermatology background, ABCD rule, transfer-learning rationale, methodology, results, clinical limitations)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, class imbalance, ABCD rule, medical-AI limits)

• This is an educational prototype, not a medical device: it cannot diagnose, rule out or grade any skin condition. Any real skin concern must be assessed by a qualified dermatologist with dermoscopy and histopathology. The report, demo and page all state this.
• ≥80% balanced accuracy and ≥85% melanoma recall are design targets for the training run, stated honestly — the report documents the actual achieved figures after training.
• The demo's ABCD score is a transparent clinical rule computed from the photo, not the CNN output — the report and demo both say this explicitly.
• HAM10000 is dermatoscopic (magnified, polarized); phone-camera close-ups differ in lighting and scale — the report states this domain gap.
• Rare classes (df, vasc) have few examples; their metrics carry wide uncertainty, reported honestly in the per-class table.