[Paper Review] Dermatologist Level Dermoscopy Skin Cancer Classification Using Different Deep Learning Convolutional Neural Networks Algorithms

Amirreza Rezvantalab, Habib Safigholi+1 more|arXiv (Cornell University)|Oct 21, 2018

Cutaneous Melanoma Detection and Management31 references100 citations

TL;DR

The paper evaluates pre-trained CNNs on eight dermoscopy skin cancer classes, showing CNNs outperform dermatologists, with DenseNet201 and ResNet152 achieving top ROC-AUC scores.

ABSTRACT

In this paper, the effectiveness and capability of convolutional neural networks have been studied in the classification of 8 skin diseases. Different pre-trained state-of-the-art architectures (DenseNet 201, ResNet 152, Inception v3, InceptionResNet v2) were used and applied on 10135 dermoscopy skin images in total (HAM10000: 10015, PH2: 120). The utilized dataset includes 8 diagnostic categories - melanoma, melanocytic nevi, basal cell carcinoma, benign keratosis, actinic keratosis and intraepithelial carcinoma, dermatofibroma, vascular lesions, and atypical nevi. The aim is to compare the ability of deep learning with the performance of highly trained dermatologists. Overall, the mean results show that all deep learning models outperformed dermatologists (at least 11%). The best ROC AUC values for melanoma and basal cell carcinoma are 94.40% (ResNet 152) and 99.30% (DenseNet 201) versus 82.26% and 88.82% of dermatologists, respectively. Also, DenseNet 201 had the highest macro and micro averaged AUC values for overall classification (98.16%, 98.79%, respectively).

Motivation & Objective

• Assess the effectiveness of pre-trained CNN architectures for classifying eight dermoscopy skin diseases.
• Compare multiple CNN architectures (DenseNet201, ResNet152, Inception v3, InceptionResNet v2) on a combined dataset.
• Quantify model performance relative to dermatologist benchmarks using ROC AUC metrics.
• Identify which architecture yields the highest macro- and micro-averaged AUC for overall classification.

Proposed method

• Use pre-trained CNN architectures DenseNet201, ResNet152, Inception v3, and InceptionResNet v2.
• Apply models to a combined dataset of 10,135 dermoscopy images (HAM10000: 10,015; PH2: 120).
• Classify images into eight diagnostic categories: melanoma, melanocytic nevi, basal cell carcinoma, benign keratosis, actinic keratosis and intraepithelial carcinoma, dermatofibroma, vascular lesions, atypical nevi.
• Evaluate performance using ROC AUC metrics for each class and overall.
• Compare CNN performance to dermatologist performance reported in the abstract.
• Report best-performing architectures for specific cancer types (e.g., melanoma, BCC).

Experimental results

Research questions

• RQ1 Can pre-trained CNNs match or exceed dermatologist performance on eight dermoscopy classes?
• RQ2 Which CNN architecture provides the highest performance (ROC AUC) for melanoma and basal cell carcinoma?
• RQ3 What are the macro- and micro-averaged AUC values achieved by top-performing models for overall classification?

Key findings

• All deep learning models outperformed dermatologists by at least 11%.
• Melanoma ROC AUC: 94.40% with ResNet152; Basal cell carcinoma ROC AUC: 99.30% with DenseNet201.
• Dermatologists' ROC AUC: 82.26% for melanoma and 88.82% for basal cell carcinoma.
• DenseNet201 achieved the highest macro- and micro-averaged AUC for overall classification (98.16% and 98.79%, respectively).