[Paper Review] Transfer Learning for Melanoma Detection: Participation in ISIC 2017 Skin Lesion Classification Challenge

Dennis H. Murphree, Che Ngufor|arXiv (Cornell University)|Mar 15, 2017

Cutaneous Melanoma Detection and Management5 references16 citations

TL;DR

This paper presents a transfer learning approach using pre-trained Inception V3 for melanoma detection in the ISIC 2017 challenge, employing feature extraction and fine-tuning on a multi-layer perceptron. The method achieved an average AUC of 0.80, with 0.84 on distinguishing seborrheic keratosis from nevus/melanoma and 0.76 on distinguishing melanoma from nevus/seborrheic keratosis.

ABSTRACT

This manuscript describes our participation in the International Skin Imaging Collaboration's 2017 Skin Lesion Analysis Towards Melanoma Detection competition. We participated in Part 3: Lesion Classification. The two stated goals of this binary image classification challenge were to distinguish between (a) melanoma and (b) nevus and seborrheic keratosis, followed by distinguishing between (a) seborrheic keratosis and (b) nevus and melanoma. We chose a deep neural network approach with a transfer learning strategy, using a pre-trained Inception V3 network as both a feature extractor to provide input for a multi-layer perceptron as well as fine-tuning an augmented Inception network. This approach yielded validation set AUC's of 0.84 on the second task and 0.76 on the first task, for an average AUC of 0.80. We joined the competition unfortunately late, and we look forward to improving on these results.

Motivation & Objective

• To address the challenge of classifying skin lesions into melanoma, nevus, and seborrheic keratosis using deep learning.
• To evaluate the effectiveness of transfer learning with pre-trained models in medical image classification with limited training data.
• To improve classification performance on two binary classification tasks: melanoma vs. benign lesions and seborrheic keratosis vs. nevus/melanoma.
• To explore the impact of data augmentation and fine-tuning on model generalization in dermatoscopic image analysis.

Proposed method

• Utilized a pre-trained Inception V3 network as a feature extractor for transfer learning.
• Employed the extracted features as input to a multi-layer perceptron for classification.
• Applied fine-tuning to an augmented Inception V3 network to adapt the model to the specific skin lesion dataset.
• Used data augmentation techniques to improve model robustness and reduce overfitting.
• Trained the model on the ISIC 2017 dataset, focusing on two binary classification tasks.
• Optimized the model using validation set AUC as the primary evaluation metric.

Experimental results

Research questions

• RQ1 Can transfer learning with a pre-trained Inception V3 model effectively improve skin lesion classification performance in a low-data regime?
• RQ2 How does fine-tuning an Inception V3 network compare to feature extraction alone in distinguishing melanoma from benign lesions?
• RQ3 What is the impact of data augmentation on model generalization in the context of dermatoscopic image classification?
• RQ4 How does the model perform on the two-tiered classification task: first distinguishing melanoma, then separating nevus from seborrheic keratosis?
• RQ5 To what extent does transfer learning mitigate overfitting in a small medical imaging dataset?

Key findings

• The model achieved a validation set AUC of 0.84 on the second classification task, distinguishing seborrheic keratosis from nevus and melanoma.
• The model achieved a validation set AUC of 0.76 on the first classification task, distinguishing melanoma from nevus and seborrheic keratosis.
• The average AUC across both tasks was 0.80, indicating strong overall performance on the binary classification challenges.
• Fine-tuning the Inception V3 model with data augmentation improved generalization and performance on the validation set.
• The transfer learning approach proved effective despite late participation in the competition, suggesting strong potential for future improvements.
• The use of a multi-layer perceptron on top of Inception V3 features provided a robust and efficient classification pipeline.