BILSK: A bilinear convolutional neural network approach for skin lesion classification

BILSK: A bilinear convolutional neural network approach for skin lesion classification

Blog Article

Background and objective: Skin lesions are areas of the dermis that have an abnormal growth or appearance compared to the surrounding area.They can be harmless such as a small scrape or severe as skin cancer.These dermis abnormalities increase in size over time and cause morbidity problems.Correct diagnosis of these skin lesions is crucial new belial model for successful treatment, which is generally too expensive.

Convolutional neural networks (CNN) have been investigated for the classification of skin lesions with different training methods and techniques.However, the best results obtained by previous works show that there is a wide range to be achieved regarding detection, precision, and computational costs.Methods: Therefore, in this paper, we present a bilinear CNN approach capable of classifying seven skin lesions classes with the highest accuracy of the state-of-the-art and low computational cost.The framework proposed la rams crop top includes a data augmentation step to correct the data imbalance problem, transfer learning, and fine-tuning to improve the classification performance while reducing the computational cost.

Several simulations were executed over the HAM10000 dataset.Results: The results show that a bilinear approach composed of the ResNet50 and the VGG16 architectures, increases accuracy by 2.7% compared to the state-of-the-art.Specifically, the proposed approach achieves an average of 0.

9321 accuracy, 0.9292 precision, 0.9300 recall, 0.9321 F1 score, 0.

9810 AUC, and requires 238.6 min for training.Conclusions: This performance increase can help to support the clinician diagnosis in order to provide a second opinion which can reduce the morbidity and treatment costs.