Basic Attention U-Net：適用於醫療影像語意分割之深度學習網路

Basic Attention U-Net: A Deep Learning Neural Networkfor Medical Image Semantic Segmentation

近年來許多研究致力於將深度學習應用於醫學領域，而深度學習在醫學影像方面也取得了相當顯著的成果。然而，醫學影像中每個病灶特徵的形狀和大小各異，可能導致分類模型在學習的過程中無法正確判斷。本研究著眼於醫學影像的影像分割，以便先行準確地定位出醫學影像中可能的病灶。本研究提出一種新的深度學習神經網路模型，採用U-Net的基本模塊做為基礎，結合了通道注意力機制（Channel Attention）以及空間注意力機制（Spatial Attention），命名為BAU-Net（Basic Attention U-Net）。該網路能夠有效地融合影像中空間及通道的資訊，使模型能夠捕捉到不同的資訊，進而取得良好的預測結果。在ISIC2018公開的資料集上的實驗結果顯示，本研究所提出的BAU-Net在mIoU（Mean Intersection over Union）指標得到0.871，表示平均上預測的分割區域與實際病灶區域重疊為87.1%，相當優秀。這表明了BAU-Net在醫學影像語意分割任務上表現相當不錯。

In recent years, numerous studies have focused on applying deep learning to the field of medicine, with significant achievements in medical imaging. However, the diverse shapes and sizes of lesion features in medical images may lead classification models to learn incorrect information during the training process. This study focuses on medical image segmentation, aiming to accurately locate lesions in medical images. The study proposes a new neural network module, named BAU-Net, which uses U-Net as basic module and combines channel attention mechanisms and spatial attention mechanisms. This network effectively integrates spatial and channel information, enabling the model to capture diverse information and achieve robust predictive results. Experimental results on the ISIC2018 dataset demonstrate that the mloU (Mean Intersection over Union) metric of BAU-Net is 0.871. This indicates the high performance of BAU-Net in medical image segmen缸tion tasks.