کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

Enhancing Precision in Dermoscopic Imaging using TransUNet and CASCADE

نویسندگان :

Mahdi Niknejad¹ Mahdi Firouzbakht² Maryam Amirmazlaghani³

1- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 2- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 3- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران)

کلمات کلیدی :

Deep Learning،Dermoscopic Imaging،Medical Image Segmentation،UNet،Vision Transformer

چکیده :

Medical imaging is advancing at a rapid pace, revolutionizing medicine. Skin lesion segmentation is essential when employing computer-aided approaches for early skin cancer diagnosis. However, automatically segmenting skin lesions in dermoscopic images is a difficult task because of several challenges. Despite recent good performance, CNN-based algorithms are unable to effectively learn explicit global and long-range semantic information because of the fundamental locality of convolution operations. The first medical image segmentation framework, TransUNet, was presented employing Vision Transformer as a strong encoder in a U-shaped architecture, in light of the growing interest in self-attention mechanisms in computer vision and their potential to address this issue. Using hierarchical vision transformers' multi-scale features, CASCADE is a novel attention-based decoder. The components of CASCADE are a convolutional attention module that improves the local and long-range context by suppressing background information and an attention gate that combines features. Using TransUNet's Encoder and CASCADE as the Decoder, the TransCASCADE model efficiently makes use of the global context stored by Transformers and detailed, high-resolution spatial information from CNN features. Experimental results demonstrate the high efficiency of the proposed method on PH2 dataset.