Effective Melanoma Recognition Using Deep Convolutional Neural Network with Covariance Discriminant Loss

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Melanoma recognition faces challenges due to imbalanced datasets, high intra-class variation, and inter-class similarity in dermoscopy images.
• Existing deep learning models struggle to effectively differentiate between melanoma and non-melanoma lesions under these conditions.

Discussion:

• The proposed method utilizes a deep convolutional neural network (CNN) trained with both cross-entropy loss and a novel covariance discriminant loss.
• Covariance discriminant loss simultaneously considers first and second-order distances, providing enhanced constraints for feature extraction.
• This loss function specifically constrains the distance between hard samples and the minority class center, improving the separation of deep features for melanoma and non-melanoma.
• An algorithm for mining hard samples is developed to further refine the model's performance.

Key Insights:

• The integration of covariance discriminant loss with CNNs significantly improves melanoma recognition accuracy.
• The method effectively handles data imbalance and feature variations inherent in skin lesion datasets.
• Achieved sensitivities of 0.942 and 0.917 on the ISBI 2018 Skin Lesion Analysis dataset demonstrate the model's efficacy.

Outlook:

• Further validation on diverse and larger datasets is warranted to confirm generalizability.
• Exploration of integrating this loss function into other medical image analysis tasks could be beneficial.
• Potential for clinical application in automated melanoma detection systems.