【深度学习】医学图像处理之视杯视盘分割调研分析

【深度学习】医学图像处理之视杯视盘分割数据集和评价指标

文章目录

• • 【深度学习】医学图像处理之视杯视盘分割数据集和评价指标

• 1 数据集（公开）

• • 2.1 视盘标签
  • 2.2 视杯视盘标签

• 2 评价指标

• • 2.1 absolute CDR (Cup to Disc Ratio) error（杯盘比错误率）
  • 2.2 其他指标

• 3 总结

1 数据集（公开）

2.1 视盘标签

ORIGA-650
这批数据集归属于新加坡国家眼科中心，主要包含650张彩色眼底图像，每张图像都有视盘和视杯的分割标注，同时还有是否患有青光眼的诊断标注。拥有这批数据的IMED团队，也是目前国内最大的眼科医疗图像组。 ORIGA-650分为两个子集 set A for training 和 set B for testing 每个子集包含了325张图像。
Messidor
Messidor数据集原本是用来做糖尿病视网膜（diabetic retinopathy， DR）病变检测的，只有糖网的分级标注。后来国外的一个课题组又重新手工标定了视盘的边界，因此目前大家也同样在Messidor数据上做视盘的定位和分割。
RIM-ONE
RIM-ONE一共发布了三个子数据集（RIM-ONE-R1，R2，R3），他们的数量分别是169，455和159张。
DRION-DB
DRION-DB 做的人特别少，但是这批数据集也有111张图像。大家也可以做一下，就当作一种data augmentation了吧。

2.2 视杯视盘标签

除上图中常用的Refuge数据集外，在最新的研究中使用较多的是GAMMA数据集。

MICCAI2021 Contest : GAMMA
数据简介Dataset Introduction

该数据集由中国广州中山大学中山眼科中心提供。该集合共包含300对两种临床眼底影像数据。这些影像分别涉及2D眼底彩照和3D OCT检查。我们将各类别数据按照三分之一的比例分配给参赛者。这些分配将用于训练阶段、预赛阶段和决赛阶段的比赛准备工作。对于深度学习算法而言，在这100对小样本训练数据中进行模型训练是必要的前提条件。我们鼓励所有参赛者就这一主题提出创新性的解决方案以提升模型性能。

The dataset released by GAMMA was made available by the Sun Yat-sen Ophthalmic Center and Sun Yat-sen University in Guangzhou, China. It includes 300 sets of two different clinical imaging techniques. These techniques include 2D fundus color photography and 3D OCT imaging, both of which are widely used in clinical fundus examinations. We categorized all data into three equal groups and allocated them for training, preliminary testing, and final evaluation processes. For the deep learning algorithm, the 100 training data pairs belong to a small number of samples; therefore, we encourage participants to propose models optimized for small sample datasets.

GAMMA 数据集中的数据包括了属于青光眼分级情况、黄斑中心凹的位置信息以及眼底图像中杯状与盘状结构的分割标记。采用三个子任务作为基础, 将详细阐述各类标注的具体实现流程。

The GAMMA dataset contained specific glaucoma grades, precise fovea coordinates, and a detailed mask of the cup and optic disc. The subsequent sections detail the implementation of various annotations based on three distinct subtasks.

各项数据的青光眼分级标准基于临床记录。该标准是基于全面且多维度的临床检查数据得出的结论。 clinical check-up data. 临床检查涵盖眼底彩照、眼压测量、OCT, 视力表测试等其他相关技术.

True outcomes in the classification of glaucoma for each dataset were established based on comprehensive clinical documentation and rely upon the aggregated results from all clinical assessments. Clinical evaluations encompass a range of procedures including color imaging of the retina (e.g., fundus photography), intra-ocular pressure assessment (e.g., tonometry), OCT scans, visual field testing (e.g., standard automated perimetry), and additional tests as required.

任务二：黄斑中央凹定位

用于确定黄斑中央凹坐标的初始数据

The initial fovea coordinate annotation process for each data set was carried out manually by four clinical ophthalmologists from Sun Yat-sen Ophthalmic Center at Sun Yat-sen University in China. These experts, with an average experience level of eight years (ranging from five to ten years), independently identified the fovea within images using a cross marker tool without access to any patient data or disease prevalence information. The findings obtained from these four ophthalmologists were then integrated by a senior ophthalmologist with over ten years of glaucoma expertise. This senior professional examined the four sets of markers and made necessary adjustments through methods such as averaging their positions or performing fine-tuning on specific points to ensure accuracy and consistency.

该任务与任务二具有相似性，在不掌握任何患者信息或数据流行病学情况下完成每个数据集的初始视杯视盘分割区域标注工作。该过程由4位临床眼科专家独立完成：首先每位专家分别对图像中视盘和视杯区域进行勾勒；随后将4个初始标注结果汇总至任务二的高级专家团队中进行融合。采用多数投票法进行融合后确定的金标准：最终确定的金标准通过取所有参与投票专家标注结果的交集确定。

Similar to Task 2, manual annotation of the optic cup and disc's initial segmentation region was carried out by four clinical ophthalmologists from the Sun Yat-sen University Zhongshan Ophthalmic Center for each data point. These experts independently performed optic disc and cup segmentation in the images without referencing patient information or disease prevalence. The four segmentation results were then reviewed by a senior ophthalmologist in Task 2 for integration. A majority vote was employed to fuse the outcomes of optic cup and optic disc segmentation. The senior fusionist examined the initial marking data and selected the overlapping regions across multiple ophthalmologists as the definitive golden standard for optic cup and disc segmentation.

数据描述Data Introduction/Description

training dataset

GAMMA launched their training dataset, comprising one hundred samples labeled from "S" to "T". Each sample corresponds to a dedicated folder structure (e.g., "Sample S-457") that includes both a two-dimensional fundus image and a three-dimensional OCT scan. Within this dataset, all samples are numbered from "S" to "T", with each containing an associated storage location. All corresponding images are organized into folders marked with their respective IDs (e.g., S-457). Each sample's storage location was created under specific imaging protocols for optimal resolution.

任务三：遵循.bmp 图像格式保存 视杯视盘分割金标准 每个样本的眼底图像是一个配对 的 分割结果 图像是吗？ 其命名与输入待处理的眼底图象名前缀相同 在此过程中 每个 分割后的 图像 中 像素值为 0 表示 视杯 区域 像素值为 128 表示 视盘 中非 视杯 区域 像素值为 255 则表示其他区域能够实现这一目标 所有 样本的所有 分割 图档 存档于.Disc_Cup_Masks 文件夹内

预赛学习者数据集Learner dataset 包括了100个样本的数据对（其中包含了100个样本的数据对）。
这些样本遵循相同的存储格式。

The dataset of the preliminary competition includes data pairs spanning from 0101 to 0200 encompassing a total of 100 samples. Furthermore, the manner in which data is stored within this dataset aligns precisely with the format utilized in the training set.

决赛数据集Final set 比赛数据集仅限于已晋级的队伍 数据集中涵盖100个样本的数量 日期范围为2月1日至3月日

The final dataset is solely intended for the final teams. It includes a total of 100 sample entries spanning from year codes 201 to 300. The data storage format strictly complies with the training set’s format.

2 评价指标

2.1 absolute CDR (Cup to Disc Ratio) error（杯盘比错误率）

CDR是衡量杯与盘之间横向尺寸、纵向尺寸以及面积的比例指标。当该比例数值达到小于或等于0.3时，则判断被评估者的眼部健康状况属于正常范围之内；然而，在实际测量中若出现单眼 cup/disk 比例超过 0.3 的情况，则应立即判定被试者存在眼底结构异常风险。根据公式（1）至（3），我们能够系统地计算并确定多种杯盘比率参数的具体数值以辅助青光眼早期筛查工作

计算方法：

2.2 其他指标

Iou计算：

在目标检测研究领域中, IOU的计算是不可或缺的重要组成部分.例如, 在R-CNN网络中, 正样本与负样本的区分通常基于候选框与真实框之间的IOU值大小.这一细节值得深入探讨并单独阐述.在R-CNN网络架构中, 我们通过Selective Search（SS）算法, 可以实现在每张图片上可以生成大约2000个候选框.如何将这些候选框划分为正样本和负样本呢?它源自于数学集合论的概念, 用于衡量两个集合的空间交叠程度.其计算公式如下：

    def calc_iou(self, boxes1, boxes2, scope='iou'):
        """calculate ious
        Args:
          boxes1: 5-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL, 4]  ====> 4：(x_center, y_center, w, h)
          （2,7,7,2,4）
          boxes2: 5-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL, 4] ===> 4：(x_center, y_center, w, h)
          （2,7,7,2,4）
        Return:
          iou: 4-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL]  --（2,7,7,2）
        """
        with tf.variable_scope(scope):
            # transform (x_center, y_center, w, h) to (x1, y1, x2, y2)
            boxes1_t = tf.stack([boxes1[..., 0] - boxes1[..., 2] / 2.0,
                                 boxes1[..., 1] - boxes1[..., 3] / 2.0,
                                 boxes1[..., 0] + boxes1[..., 2] / 2.0,
                                 boxes1[..., 1] + boxes1[..., 3] / 2.0],
                                axis=-1)  #tf.stack：矩阵拼接
    
            boxes2_t = tf.stack([boxes2[..., 0] - boxes2[..., 2] / 2.0,
                                 boxes2[..., 1] - boxes2[..., 3] / 2.0,
                                 boxes2[..., 0] + boxes2[..., 2] / 2.0,
                                 boxes2[..., 1] + boxes2[..., 3] / 2.0],
                                axis=-1)
    
            # calculate the left up point & right down point
            lu = tf.maximum(boxes1_t[..., :2], boxes2_t[..., :2]) #左上角坐标最大值
            rd = tf.minimum(boxes1_t[..., 2:], boxes2_t[..., 2:]) #右下角坐标最小值
    
            # intersection
            intersection = tf.maximum(0.0, rd - lu)
            inter_square = intersection[..., 0] * intersection[..., 1]
    
            # calculate the boxs1 square and boxs2 square
            square1 = boxes1[..., 2] * boxes1[..., 3]
            square2 = boxes2[..., 2] * boxes2[..., 3]
    
            union_square = tf.maximum(square1 + square2 - inter_square, 1e-10)
    
        return tf.clip_by_value(inter_square / union_square, 0.0, 1.0) #截断操作，即如果值不在指定的范围里，那么就会进行最大小值截断

3 总结

研究不足
1.研究人员经常评估特征的数量以进行分类,但仍有一些特征值得进一步探索以优化分类器性能[
2.在视网膜中存在低对比度且圆盘与杯子之间不可见边界的情况下,大的和小尺寸视盘及视杯的分割或过度分割被视为挑战性问题。
3.随着图像数量的增长而提高分类精度的努力未取得显著成果。
4.由于存在视网膜血管及周围萎缩的情况,分割性能受到影响。
5.对视杯区域划分方面的工作相对较少,导致相关诊断指标未能得到充分评估。
6.随着复杂性增加,现有的诊断参数未能全面综合考虑所有影响因素,限制了分析效果。
7.尽管复杂性有所提升,但现有方法仍需进一步完善以提高分割效率与准确性之间的平衡点。
8.现有分割与分类方法缺乏针对大规模数据集临床验证的支持机制,影响其临床应用效果。

发现
1.相较于其他通道而言,红色通道更适合于视盘分割,这是因为与其他通道相比,该通道中的视盘具有清晰的外观.
2.由于该通道具有显著的对比度,与其他通道相比,绿色通道更适合于视杯的分割.
3.形态学方法和各种滤波器用于从视网膜图像中去除血管,这可能导致分割难度增加.
4.从讨论中可以看出,每种方法都有各自的优缺点.
5.表4中列出的优点和缺点取决于一些因素.
6.可以分析这种分类器对分类问题并没有最佳解决方案.
7.每种技术都存在其自身的优缺点,如表5所示.
8.这些技术的具体表现取决于输入数量、输入质量和分类时间等因素.
9.然而,在某些情况下,SVM分类器因其高精度而被广泛采用.

除了第6.2节所述的结果之外，在对数据集中的现有技术方法进行性能评估时