NBAmine/ham10000-benchmarks

HAM10000 Skin Lesion Classification

Model Description

This model is a fine-tuned version of a computer vision architecture (YOLO, Transformer, or CNN) trained on the HAM10000 dataset. The primary objective is the binary classification of skin lesions into two categories: Benign and Malignant. This work is part of a benchmark study evaluating modern architectures for dermatological triage.

• Task: Binary Image Classification
• Dataset: HAM10000 (Human Against Machine)
• Classes: Benign, Malignant
• Developer: NBAmine

Objective & Methodology

The goal of this project is to identify the most effective architecture for early skin cancer detection. We evaluate several models including:

• YOLO Series: v5n, v8n, 11n (Classification)
• Transformers: ViT-B/16, Swin Tiny
• CNNs: ResNet18, ConvNeXt Tiny

Training Highlights

• Input Resolution: 224x224 pixels.
• Augmentation: 180° rotations and flips to ensure rotation invariance.
• Class Balancing: Strategic oversampling of the Malignant class to mitigate dataset imbalance and improve clinical reliability.

Evaluation Metrics

The models are evaluated with a focus on sensitivity to ensure high detection rates for malignant cases:

• Primary Metrics: Recall (Malignant class), Macro F1-Score.
• Secondary Metrics: Accuracy, Precision, Validation Loss.

Intended Use

This model is intended for research and educational purposes only. It is designed as a decision-support tool for medical imaging analysis and is not a replacement for professional clinical diagnosis by a dermatologist.