Medical Image Segmentation Based On Deep Learning

Medical image analysis plays a vital role in facilitating pathological assessment,reliable diagnosis ensuring and disease progression monitoring.As a research upsurge,deep-learning models have been widely applied in automating medical image segmentation to reduce human effort and improve the efficiency and accuracy of medical image segmentation.In this thesis,we propose a novel approach based on interventricular septum detection to generate the starting spot in an existing algorithm for left ventricular contour extraction via deep reinforcement learning.Unlike the original algorithm,which relies on supervised learning to generate the landing spot,our proposed method eliminates the need for a large dataset and time-consuming data preprocessing.By combining discrete wavelet transform with edge detection algorithm,the proposed method can accurately and effectively extract the core contour of the cardiac image,and use dynamic time warping to judge the similarity of arcs to find the target interventricular septum and generate the landing spot,which is the starting position on cardiac image coordinates.In addition,this thesis focuses on medical image segmentation based on interpretable deep learning,and proposes an interpretability metric as part of the loss function.The proposed interpretability metric comes from a Class Activation Map(CAM)with learnable weights,such that the model can be optimized to achieve an ideal balance between segmentation performance and interpretability.After feature extraction using the U-Net model,the heatmap is obtained by weighting each channel of the last feature map,so as to make the deep neural network interpretable.With the corresponding term introduced in loss function,the proposed model not only affects the interpretability module,but also improves the segmentation performance of the main segmentation network.The cardiac MRI dataset and skin cancer dataset are utilized in this paper to train and test the proposed model.Our results demonstrate that the proposed method outperforms other algorithms in terms of segmentation accuracy and overall performance.