A Review on Few-shot Learning for Medical Image Segmentation

Deep-learning based approach has solved various medical imaging problems successfully. Since the lack of training data issues caused by patient privacy, the few-shot learning method has been studied widely. However, this issue still afflicts model performance even in few-shot learning methods. To solve this issue, it is important to quickly optimize the initial parameter values using a small amount of data. In addition, to utilize small data effectively, it is important to design the objective function for segmentation suitable for GT (Ground Truth) with few-shots. In this paper, we experiment with various algorithms using the MAML (Model Agnostic Meta-Learning) method. And we propose an optimal few-shot semantic segmentation network. The proposed method uses a gradient descent algorithm and optimizer parameter decomposition method to ensure fast convergence with fewer data. Experimental results show high performance and fast convergence using fewer datasets than conventional methods.

I. Introduction

Sufficient medical data is necessary for medical advancement. However, medical data conflicts with the aspect of privacy and the ethical aspect of research. Restriction of information from the patient’s ethical point of view is a factor that adversely affects not only medical research but also the development of the medical industry. In addition, expertise in the medical field requires judgment through the doctor’s knowledge. These problems make it difficult to handle special-ized data in deep learning research related to disease or image reading, and labeled data for learning are insufficient. Recent deep learning research to solve these problems aims to create a model that can be learned with a small amount of medical data. Also, many works are being conducted. In particular, U-Net [5], SegNet [4] models have improved the performance of semantic segmentation. Meta-learning is a representative method for learning with only a small amount of data. Self-supervised meta-learning [1] (SSL-ALPNet) performs semantic segmentation through super-pixels on unannotated data sets. The MAML algorithm is the basis of our method [3]. We aim to find an objective function that optimizes the meta initial parameter θ which can be quickly applied to an unknown task. The training data, denoted as D, used for meta-learning include support and query sets. The loss function with the meta-initial parameter θ is expressed as L(θ), and task-specific parameters are then calculated as follows: \begin{equation*}{\theta ^\prime } = {\mathbb{E}_{D\sim p(D)}}\left[ {\left( {\theta - \alpha \nabla {L_D}\left( {{f_\theta }} \right)} \right)} \right],\tag{1}\end{equation*} where α represents the learning rate, and p(D) the distribution of each task’s training data. Meta-initial parameter is updated using loss which calculated by task-specific adapted parameters with gradient descent as: \begin{equation*}\theta \leftarrow \theta - \beta {\nabla _\theta }\sum\limits_{D\sim p(D)} {{L_{{D_i}}}} \left( {{f_{\theta^{\prime} i}}} \right),\tag{2}\end{equation*} where β represents the meta step size. Meta-learning is well known as ‘learning how to learn’. Since meta-learning teaches the learning process to solve a problem, it can learn quickly with a small amount of data. In addition, it solves the problem of lack of medical data. In this paper, we use a gradient descent based learning method (MAML). In addition, in terms of the loss function, experiment by adding baseline cross-entropy loss, boundary loss, and Hausdorff loss. We propose a method for improving the efficient performance of semantic segmentation in each applied algorithm. The proposed method performs better than the baseline method in the segmentation of unseen medical image data.