AI System Spots Skin Cancer With 94% Accuracy, Outpacing Human Dermatologists
Researchers have developed a machine-learning model that diagnoses melanoma and other skin diseases from images with near-expert accuracy, potentially reducing diagnostic delays and costs. The breakthrough combines dual attention mechanisms with neural networks to identify lesions doctors might miss, signaling a path toward automated screening in clinics and telemedicine platforms.
Originaltitel: An attention guided bilinear convolution neural network using hybrid loss function for efficient skin disease classification
<p>accurate diagnosis of skin diseases, particularly melanoma is crucial for effective treatment. Traditional approaches for skin disease classification face challenges due to the diverse and subjective nature of skin lesion, therefore, it is difficult to distinguish between situations that have visual similarities. This paper presents a novel deep learning based model to address these challenges. The model integrates dual attention mechanisms with a bilinear Convolution Neural Network (CNN) to extract discriminative features from dermoscopic images. Generate an Attention Heat Maps (AHMs) for the crop regions of the input images to focus on significant regions and a novel loss function, which combined complementary entropy and cross-entropy for mitigating class imbalance issues and improve model performance. The proposed model achieving an accuracy of 94.24% on ISIC2019 and 92.86% on HAM10000.</p>