Multi-task Algorithm Based On Transfer Learning For Medical Image Segmentation

Medical image segmentation is a critical step in pathology assessment and monitoring and plays a key role in the field of medical image processing.At present,deep learning technology has promoted the vigorous development of the field of medical image processing.Extensive methods tend to utilize a deep convolutional neural network for various medical image segmentation tasks.Challenges in achieving high-precision medical image segmentation arise from the inherent difficulty of medical images.The inherent segmentation problems include:the difficulty of the region boundary to be segmented and the lack of training datasets.To solve the difficult problem of segmenting the boundaries,a multi-task learning strategy combining edge detection and image segmentation is proposed.To solve the problem of lack of medical image data,a transfer learning strategy is proposed to transfer knowledge from natural image data to medical image segmentation.Therefore,according to the characteristics of medical images,we propose a multi-task algorithm for medical image segmentation based on transfer learning.The main work is as follows:Firstly,we propose a multi-task learning strategy for joint edge detection and image segmentation.Edge detection and image segmentation are two different but related visual tasks.There is a logical relationship between them,that is,edge detection can be understood as a subset of image segmentation tasks.A multi-task learning strategy can be used to combine these two tasks to implement an end-to-end edge-detectionimage-segmentation joint network to solve the above problems.The selective intersection integration module proposed in this paper deeply integrates segmentation information with edge information and finally generates high-precision segmentation result.On this basis,to make full use of the characteristics of medical images,we design a central decoupling mechanism.By decoupling the segmentation labels into detailed edge labels and segmentation labels,a multi-supervised multi-task network structure is realized.At the same time,we also propose a novel cross fusion mechanism to deeply fuse the dual-stream information in the network,that is,detail edge information and segmentation information.Experiments show that,compared with previous single-task learning strategies,multi-task learning strategies can greatly improve the accuracy of segmentation algorithms.Secondly,we investigate relation-based transfer learning strategies.Due to the small scale of public medical image datasets and few comparison methods,it is difficult to demonstrate the effectiveness of the segmentation algorithm in this paper under very fair conditions.Thus,we use a relational-based transfer learning strategy to first compare segmentation algorithm on a larger dataset of natural images,and then transfer relevant experience to medical images.Experiments show that our method significantly outperforms previous methods in both natural image segmentation and medical image segmentation tasks,proving the advantages of this strategy.Thirdly,we investigate parameter-based transfer learning strategies.Due to the lack of medical training datasets,it is difficult to solve the data starvation problem.Thus,we adopt a parameter-based transfer learning strategy.We utilize Swin Transformer as backbone,and pre-train it on the ultra-large-scale natural image dataset ImageNet,and then load the pre-trained weight for further fine-tuning on the medical image dataset.Experiments show that the multi-task algorithm for medical image segmentation based on transfer learning has achieved excellent results on four datasets of polyp segmentation and optic disc segmentation in terms of four key evaluation metrics such as Dice and IoU.