Evolutionary Computing Application to Biomedical Image Segmentation

我们的目标是开发一个进化计算平台的原型，它接受 输入特定类型的图像的随机样本，以及它们对应的基本事实。它会 输出一个计算机程序，该程序能够对相同类型的未见图像进行高度分割 分割的精确度和计算效率。该平台本身可以被描述为一种基因 一种规划方法，其对应于候选者进化节点的前馈网络的群体 图像分割程序。这种特殊的表现形式是受笛卡尔遗传的启发 编程(CGP)。每个节点由其识别号(ID)、图像变换 转换参数和提供其输入的节点的ID(S)。这种转变来自于一个 开放的转换库，其中包括众所周知的图像预处理、分割和 后处理方法。个体的适合度将从基于像素的和基于 基于区域的视角。整个进化周期搜索所有(小型)前馈的空间 具有不同结构和参数的网络，用于具有最高相对适应值的网络。 我们工作的最终价值在于它试图模仿和超越高薪工作的方式 专业人士开发用于图像分割的分布式程序。因此，我们会认为我们的努力是 只有在生产出功能齐全和可演示的原型的情况下才能取得成功。

Our objective is the development of a prototype of an evolutionary computing platform, which accepts as input a random sample of images of a certain type, along with their corresponding ground truths. It would output a computer program that is capable of segmenting unseen images of the same type, with a high degree of segmentation accuracy and computational efficiency. The platform itself could be described as a genetic programming method that evolves a population of feed-forward networks of nodes, corresponding to candidate image segmentation programs. The particular form of representation is inspired by Cartesian Genetic Programming (CGP). Each node is defined by its identification number (ID), image transformation, transformation parameters and the IDs of the nodes that provide its input(s). The transformation comes from an open library of transformations, which include well-known image pre-processing, segmentation and post-processing methods. The fitness of an individual will be measured from both a pixel-based and a region-based perspective. The overall evolutionary cycle searches the space of all (small-sized) feed-forward networks, of different structures and parameterizations, for those networks with highest relative fitness values. The ultimate value of our work lies in its attempt to emulate and exceed the manner in which highly-paid professionals develop distributed programs for image segmentation. Hence, we would consider our effort a success only whence a fully-functional and demonstrable prototype is produced.