Key Problems Of Deep Learning Techniques In Medical Image Segmentation And Classification

In order to relieve the Chinese shortage of medical resources,and meet residents’growing medical needs,the studies of medical image aided diagnosis in deep learning has become a significant field in smart medicine.To better solve the difficulties in medical images,such as multimodal imaging,multi-time-points acquisition and spatial heterogeneity,we develop several researches of segmentation and classification models for medical images using deep learning techniques with clinical guidance.The researching details are shown as following:We firstly present an automatic deep-learning based framework to realize the evaluation of liver cancer ablation.The framework is comprised by:1.the segmentation of liver,liver tumors and ablation zones in multi-phases CT scans;2.the registration;3.evaluation of ablation treatment.On the independent test-set,the proposed method achieved a Dice similarity coefficient(Dice)of 95%for liver segmentation on multi-phases CT scans,respectively;a Dice of 72.27%for liver tumors on pre-treatment portal phase CT scans,a Dice of 89.06%for ablation zones on post-treatment portal phase CT scans.The results demonstrate the framework has a strong potential to evaluate treatment of ablation for liver cancer patients.We secondly present an automatic post-processing module to refine the segmentation of deep networks.The label assignment generative adversarial network(LAGAN)is improved from the generative adversarial network(GAN)and assigns labels to the probabilistic maps of deep networks.We apply the LAGAN to segment colorectal tumors in CT scans and explore the performances of different combinations between deep networks.The LAGAN increases the Dice of FCN-32s from 81.83%to 90.82%.In the Unet-based segmentation,the LAGAN enhances the Dice from 86.67%to 91.54%.The results demonstrate that the LAGAN is a robust and flexible module,which can be used to refine the segmentation of diverse deep networks for colorectal tumors.To promote the interpretability of classification networks in deep learning,we finally propose a transparency-guided ensemble convolutional neural network(CNN)to automatically discriminate pseudoprogression(PsP)and true tumor progression(TTP)in Diffusion tensor imaging(DTI).First,three typical CNNs are trained to distinguish PsP and TTP.Subsequently,we used class-specific gradient information from convolutional layers to highlight the important regions in DTI scans.And radiologists select the most lesion-relevant layer for each CNN.Finally,the selected layers are utilized to guide the construction of a multi-scale ensemble CNN whose classification accuracy reached 90.20%,and whose specificity is promoted 20%than that of any single CNN.The results demonstrate the presented network can promote the reliability and accuracy of CNNs.In conclusion,this study specifically improves the existing deep learning models in segmentation and classification for different applications and data types.The proposed researching thoughts and frameworks are approximately flexible and could provide references for more diagnosis-aided models of medical images in the future.