[Paper Review] Deep-CLASS at ISIC Machine Learning Challenge 2018

Sara Nasiri, Matthias Jung+2 more|arXiv (Cornell University)|Jul 24, 2018

Cutaneous Melanoma Detection and Management2 references3 citations

TL;DR

This paper presents Deep-CLASS, a 19-layer convolutional neural network (CNN) framework for classifying seven skin lesion types from dermoscopic images in the ISIC 2018 challenge. Using data augmentation and fine-tuned CNNs in Caffe with MATLAB, the model achieved up to 98% accuracy for Dermatofibroma and Vascular lesions, with 82% accuracy for Melanoma, demonstrating strong performance in multi-class skin disease classification.

ABSTRACT

This paper reports the method and evaluation results of MedAusbild team for ISIC challenge task. Since early 2017, our team has worked on melanoma classification [1][6], and has employed deep learning since beginning of 2018 [7]. Deep learning helps researchers absolutely to treat and detect diseases by analyzing medical data (e.g., medical images). One of the representative models among the various deep-learning models is a convolutional neural network (CNN). Although our team has an experience with segmentation and classification of benign and malignant skin-lesions, we have participated in the task 3 of ISIC Challenge 2018 for classification of seven skin diseases, explained in this paper.

Motivation & Objective

• To develop a deep learning-based system for accurate classification of seven skin lesion types from dermoscopic images.
• To improve melanoma and other skin cancer detection through end-to-end deep learning on medical imaging data.
• To address the challenge of imbalanced datasets in skin lesion classification using data augmentation and model optimization.
• To evaluate the performance of a custom CNN framework on the ISIC 2018 benchmark dataset for disease classification.

Proposed method

• A 19-layer deep CNN architecture was designed with 11 convolutional layers, 5 max-pooling layers, and 3 fully connected layers.
• Input images were resized to 128×128×3 and augmented via rotation and flipping to increase training diversity.
• The model was implemented using the Caffe deep learning framework and trained with fine-tuned hyperparameters for optimal performance.
• MATLAB was used for system development, including image processing, parallel computing, and application deployment via App Designer.
• Class weights and loss optimization were applied to handle class imbalance in the dataset.
• Evaluation was performed on a held-out test set of 161 images from the HAM10000 dataset.

Experimental results

Research questions

• RQ1 Can a deep CNN model achieve high accuracy in classifying seven distinct skin lesion types from dermoscopic images?
• RQ2 How does data augmentation affect model generalization and performance on imbalanced skin lesion datasets?
• RQ3 What is the performance of a fine-tuned CNN model on the ISIC 2018 Task 3 benchmark for multi-class skin disease classification?
• RQ4 How do precision, recall, specificity, and F-measure vary across different skin lesion types in a real-world clinical dataset?

Key findings

• The model achieved 98% accuracy for Dermatofibroma (D) and 98% for Vascular lesions (VL), indicating strong performance on rare classes.
• Melanoma (M) classification achieved 82% accuracy, with a sensitivity (recall) of 47% and precision of 30%, highlighting challenges in detecting malignant cases.
• Melanocytic nevus (N) showed the highest F-measure (78%) and precision (90%), indicating strong model confidence and performance for benign lesions.
• The confusion matrix revealed that Melanoma was frequently misclassified as Melanocytic nevus (N), suggesting class similarity issues.
• Basal cell carcinoma (BCC) achieved 92% accuracy, with a sensitivity of 52% and specificity of 94%, indicating moderate detection capability.
• The model demonstrated robustness for benign and vascular lesions but struggled with Melanoma and AK, likely due to class imbalance and visual similarity.