Cost-efficient Deep Learning For Medical Image Analysis

Deep Learning-based methods is widely used in computer-aided medical image analysis such as lesion detection,segmentation,and classification.However,the effectiveness and scalability of Deep Learning is significantly limited by the unmet cost prerequisites of largescale annotated training samples and expensive devices.In this paper,cost-efficient models are proposed to be deployed to less expensive devices,and the performance can be improved with less annotation cost.To this end,this paper addresses several specific problems including:(1)Improving deploy-ability of 3-dimensional medical image segmentation methods by reducing its storage footprint without harming the accuracy;(2)Applying Mixed Supervised Learning(MSL)to solve the challenge of Mis-located Supporting Region problem where the spatial attention does not match the lesion location,and to guide Weakly Supervised Learning(WSL)with low-cost image-level labels;(3)Automatically selecting necessary samples which are prone to the Mis-located Supporting Region problem to conduct fine-grained annotations;(4)Alleviating MSL's dependency on expensive fine-grained annotations by training a WSL-only model with better attention accuracy and less impact from the Mis-located Supporting Region problem as the baseline.This paper aims to tackle the cost efficiency problem by addressing bottlenecks of both computational power and annotation resources.The contributions are listed below:(1)This paper proposes a non-symmetric fully convolutional segmentation network for3-dimensional medical image segmentation to reduce model footprint,where existing methods with symmetric Encoder-Decoder design take too much storage,and the proposed method rebalances the resource by making the Decoder localized to Region-of-Interest.This improvement makes end-to-end full-volume segmentation possible without compromising feature richness.In addition to remarkable speedup,the accuracy of end-to-end full volume segmentation is significantly improved by enlarging effective receptive field and capturing long-range dependency.Instead of requiring expensive workstations at clinical scenes,the network can be deployed to less expensive hardware and work at acceptable speed.(2)This paper proposes an Adversarial-Learning-based Mixed Supervised Learning(MSL)method to utilize fine-grained annotations for spatial attention rectification to resolve the Mislocated Supporting Region problem.A Contribution-blocked Global Pooling layer is designed to exclude certain regions' contribution to the classification score.By excluding the contribution of lesion regions specified by fine-grained annotations,adversarial samples are generated to guide the attention to the lesion regions.By putting all available annotation resources,the MSL method achieves significantly better accuracy at equal annotation cost.By relaxing the annotation fineness from pixel-level to bounding-box-level,the annotation cost is significantly reduced compared to existing MSL methods.(3)This paper proposes an Attention-Uncertainty-based Active Learning(AL)method to query fine-grained annotation for MSL.As the Mis-located Supporting Regions problem harms the effectiveness of existing AL metrics,the AL method proposes to evaluate attention uncertainty to effectively evaluate potential benefit of annotating a sample.Furthermore,the AL method proposes to extract features from vague attention maps for more discriminative uncertainty evaluation.The AL method significantly reduced the amount of fine-grained annotation to achieve the target accuracy.(4)This paper proposes an Attention Guided Self-Paced Learning(AG-SPL)method that follows easy-to-hard human learning pattern to improve the correctness of spatial attention.By digging into the root of the Mis-located Supporting Regions problem,the AG-SPL method proposes to use the attention uncertainty for difficulty evaluation.Additionally,by a Certain Attention Solidification strategy,pseudo-fine-grained annotations are generated from easy samples to alleviate the attention-shift caused by involving hard samples.This method outperforms existing Self-paced Learning strategies by emphasizing the Mis-located Supporting Region problem to improve attention correctness.The resulting WSL model forms a better baseline for MSL framework to achieve better accuracy with less fine-grained annotations.