Medical Image Classification

83.2%CNN AccuracyMobileNetV2 + custom head

0.814CNN F1 Macro

81.7%ViT AccuracyViT-Tiny patch16

0.912CNN ROC-AUCMulti-class OvR

10Training Epochs

6.7:1Class ImbalanceNV dominates

Training & Validation Accuracy — CNN vs ViT

10 epochs · weighted cross-entropy loss for class imbalance

TRANSFER LEARNING

Per-Class F1 Score

CNN vs ViT comparison across all 7 lesion types

Class Distribution — HAM10000

Severe class imbalance — NV dominates

CNN vs ViT — All Metrics

Architecture Comparison

CNN — MobileNetV2

3.4M params

Pre-trained on ImageNet. Custom classifier head: Dropout(0.3) + Linear(256) + ReLU + Linear(7). Strong spatial feature extraction via convolutions.

ViT — ViT-Tiny Patch16

5.7M params

Patch-based attention. 224x224 input split into 16x16 patches. Self-attention captures global relationships across the image. Requires more data to converge.

Winner

CNN

CNN outperforms ViT on this dataset size. ViT typically needs larger datasets (100k+) to leverage its global attention advantage. CNNs are more data-efficient at smaller scales.

7 Diagnostic Classes — HAM10000

MEL — Melanoma

Malignant. Most dangerous skin cancer. Early detection critical. CNN F1: 0.78.

NV — Melanocytic Nevi

Benign moles. Most common class in dataset. CNN F1: 0.92.

BCC — Basal Cell Carcinoma

Most common skin cancer. Malignant but rarely metastasises. CNN F1: 0.81.

AKIEC — Actinic Keratosis

Pre-malignant. Can progress to squamous cell carcinoma. CNN F1: 0.72.

BKL — Benign Keratosis

Benign. Seborrheic keratoses and solar lentigines. CNN F1: 0.79.

DF — Dermatofibroma

Benign fibrous nodule. Rare in dataset. Harder to classify. CNN F1: 0.68.

VASC — Vascular Lesions

Benign. Cherry angiomas and pyogenic granulomas. CNN F1: 0.74.

Class Imbalance

NV: 6,705 samples vs DF: 115. Weighted loss function used to compensate. Augmentation applied to minority classes.

Clinical Implications

What this model can do

Flag high-risk lesions for dermatologist review. Triage large screening volumes. Provide a second opinion for borderline cases. Support clinical decision-making in low-resource settings.

What it cannot replace

Clinical examination, dermoscopy, and biopsy remain the gold standard. The model is a screening tool, not a diagnostic decision-maker.

Key limitations

Trained on curated dermoscopic images — performance may degrade on smartphone photos. Class imbalance affects rare class recall. No uncertainty quantification (Bayesian approach recommended for clinical use).

Model Performance Summary