[Paper Review] A Dense CNN approach for skin lesion classification

Pierluigi Carcagnì, Andrea Cuna+1 more|arXiv (Cornell University)|Jul 17, 2018

Cutaneous Melanoma Detection and Management3 references4 citations

TL;DR

This paper proposes a 61-layer DenseNet-based CNN with a center loss-regularized training strategy for classifying seven types of skin lesions from dermoscopic images. By fine-tuning a pre-trained DenseNet-121 on the ISIC 2018 dataset using a combined softmax and center loss function (λ = 0.8), the model achieves 89.2% validation accuracy, demonstrating improved feature discrimination and strong performance in melanoma and benign lesion classification.

ABSTRACT

This article presents a Deep CNN, based on the DenseNet architecture jointly with a highly discriminating learning methodology, in order to classify seven kinds of skin lesions: Melanoma, Melanocytic nevus, Basal cell carcinoma, Actinic keratosis / Bowen's disease, Benign keratosis, Dermatofibroma, Vascular lesion. In particular a 61 layers DenseNet, pre-trained on IMAGENET dataset, has been fine-tuned on ISIC 2018 Task 3 Challenge Dataset exploiting a Center Loss function.

Motivation & Objective

• To develop a deep learning model for automated classification of seven skin lesion types in dermoscopic images.
• To improve feature discriminability in skin lesion classification using a center loss regularization technique.
• To achieve high performance on the ISIC 2018 Challenge Task 3 dataset without extensive pre-processing.
• To evaluate the effectiveness of combining softmax and center loss in a fine-tuned DenseNet architecture.
• To explore the impact of data augmentation and training strategy on model generalization in imbalanced skin lesion datasets.

Proposed method

• A 61-layer DenseNet-121 variant was constructed by reducing the number of dense blocks from 4 to 3 and truncating after the third dense block, resulting in a 22-layer final dense block instead of 48.
• The network was initialized with ImageNet-pretrained weights and fine-tuned on the ISIC 2018 Challenge dataset using stochastic gradient descent (SGD) with a base learning rate of 0.01, weight decay of 0.0001, and momentum of 0.9.
• A hybrid loss function was employed, combining cross-entropy softmax loss (L_s) and center loss (L_c), with a balancing hyperparameter λ = 0.8 to optimize intra-class compactness and inter-class separation.
• Data augmentation was applied via geometric transformations (rotation, flipping, affine) to balance class distributions and increase training samples.
• Input images were cropped to a square and resized to 224×224 pixels to match the network’s input requirement.
• The final fully connected layer was trained from scratch while all earlier layers were frozen during fine-tuning.

Experimental results

Research questions

• RQ1 Can a DenseNet-based CNN with center loss regularization improve classification accuracy for diverse skin lesion types?
• RQ2 How does the combination of softmax and center loss affect feature clustering and inter-class separation in skin lesion classification?
• RQ3 To what extent does data augmentation improve performance on an imbalanced skin lesion dataset?
• RQ4 Does fine-tuning a pre-trained DenseNet on dermoscopic images yield high performance without complex pre-processing?
• RQ5 What is the optimal balance between softmax and center loss for maximizing validation accuracy in skin lesion diagnosis?

Key findings

• The proposed model achieved an overall validation accuracy of 89.2% on the ISIC 2018 Task 3 dataset, demonstrating strong performance across all lesion classes.
• The use of center loss with λ = 0.8 significantly improved intra-class compactness, enhancing model discriminability.
• After data augmentation, the training set contained 295,375 images and the test set 74,004 images, with balanced class distributions across all seven lesion types.
• The model achieved high performance with minimal pre-processing, relying only on image resizing and augmentation.
• The final network structure, with 61 layers and three dense blocks, was effective in capturing hierarchical features from dermoscopic images.
• The results suggest that center loss regularization is particularly effective in medical image classification tasks with limited and imbalanced data.