ErdanC / Tooth-and-alveolar-bone-segmentation-from-CBCT

205 stars 30 forks source link

About the 3rd stage(individual tooth segmentation) #10

Open Tony2408 opened 1 year ago

Tony2408 commented 1 year ago

第三阶段单颗牙齿的语义分割中,在自己的数据上遇到了一个问题想请教一下。 我这个实验没有沿用第一阶段和二阶段的结果,而是直接在mask上提取了准确的centroid,生成了对应的patch。数据首先baseline是单通道输入patch,测试集的dice、hd、asd就都基本已经达到了您实验的sota;尝试concat了骨架(因为考虑到传播误差,这个骨架甚至不是预测的骨架,而是直接通过gt得到的),结果是测试集dice反而下降了0.3个点,hd持平,asd小幅下降;加入multi-task的edge以后也和concat skl的情况一样,没有明显的提升。 试验后我分析,我这种做法是完全建立在centroid完全准确估计这一前提条件下的。也看到您之前的回答提到了实际上您的实验centroid加入了扰动,不会有绝对准确的可能。 我的问题在于,可否得出如下结论:倘若centroid能够无限接近其真实位置,那我们也能够在不加skl和edge、landmark的情况下使得单牙分割达到sota

lhnows commented 1 year ago

第三阶段单颗牙齿的语义分割中,在自己的数据上遇到了一个问题想请教一下。 我这个实验没有沿用第一阶段和二阶段的结果,而是直接在mask上提取了准确的centroid,生成了对应的patch。数据首先baseline是单通道输入patch,测试集的dice、hd、asd就都基本已经达到了您实验的sota;尝试concat了骨架(因为考虑到传播误差,这个骨架甚至不是预测的骨架,而是直接通过gt得到的),结果是测试集dice反而下降了0.3个点,hd持平,asd小幅下降;加入multi-task的edge以后也和concat skl的情况一样,没有明显的提升。 试验后我分析,我这种做法是完全建立在centroid完全准确估计这一前提条件下的。也看到您之前的回答提到了实际上您的实验centroid加入了扰动,不会有绝对准确的可能。 我的问题在于,可否得出如下结论:倘若centroid能够无限接近其真实位置,那我们也能够在不加skl和edge、landmark的情况下使得单牙分割达到sota

我是按照您这个思路做的,测试结果牙根稍微短了一点

Tianshuiyue commented 1 year ago

第三阶段单颗牙齿的语义分割中,在自己的数据上遇到了一个问题想请教一下。 我这个实验没有沿用第一阶段和二阶段的结果,而是直接在mask上提取了准确的centroid,生成了对应的patch。数据首先baseline是单通道输入patch,测试集的dice、hd、asd就都基本已经达到了您实验的sota;尝试concat了骨架(因为考虑到传播误差,这个骨架甚至不是预测的骨架,而是直接通过gt得到的),结果是测试集dice反而下降了0.3个点,hd持平,asd小幅下降;加入multi-task的edge以后也和concat skl的情况一样,没有明显的提升。 试验后我分析,我这种做法是完全建立在centroid完全准确估计这一前提条件下的。也看到您之前的回答提到了实际上您的实验centroid加入了扰动,不会有绝对准确的可能。 我的问题在于,可否得出如下结论:倘若centroid能够无限接近其真实位置,那我们也能够在不加skl和edge、landmark的情况下使得单牙分割达到sota

您好,请问您是成功复现了这个项目吗?在single tooth segmentation这一阶段,需要{'image': image, 'label': label, 'centroid': centroid, 'boundary': boundary, 'keypoints': keypoints}这些进行训练,请问boundary, keypoint您是如何得到的呢?

dangis-lll commented 1 year ago

您好,请问您是如何直接在mask上提取了准确的centroid的 @Tony2408

woailunhua commented 1 year ago

第三阶段单颗牙齿的语义分割中,在自己的数据上遇到了一个问题想请教一下。 我这个实验没有沿用第一阶段和二阶段的结果,而是直接在mask上提取了准确的centroid,生成了对应的patch。数据首先baseline是单通道输入patch,测试集的dice、hd、asd就都基本已经达到了您实验的sota;尝试concat了骨架(因为考虑到传播误差,这个骨架甚至不是预测的骨架,而是直接通过gt得到的),结果是测试集dice反而下降了0.3个点,hd持平,asd小幅下降;加入multi-task的edge以后也和concat skl的情况一样,没有明显的提升。 试验后我分析,我这种做法是完全建立在centroid完全准确估计这一前提条件下的。也看到您之前的回答提到了实际上您的实验centroid加入了扰动,不会有绝对准确的可能。 我的问题在于,可否得出如下结论:倘若centroid能够无限接近其真实位置,那我们也能够在不加skl和edge、landmark的情况下使得单牙分割达到sota

您好,请问您是成功复现了这个项目吗?在single tooth segmentation这一阶段,需要{'image': image, 'label': label, 'centroid': centroid, 'boundary': boundary, 'keypoints': keypoints}这些进行训练,请问boundary, keypoint您是如何得到的呢? 他连牙齿质心都是自己做的。你问牙根和牙齿边界怎么来的,你说呢- -

koushiyu commented 1 year ago

有mask啥都能算出来

dangis-lll commented 1 year ago

有mask啥都能算出来

您用于提取centroid的mask是单颗牙的吗,还是论文里的那样的所有牙列在一起的tooth roi

koushiyu commented 1 year ago

有mask啥都能算出来

您用于提取centroid的mask是单颗牙的吗,还是论文里的那样的所有牙列在一起的tooth roi

image

右上角的结果直接由mask计算得到,而不是网络的输出

dangis-lll commented 1 year ago

snapshot0001 就是直接通过这样的toothroi,利用非深度学习方法就能获得这些中心是吗,大佬方便详细说说吗,或者指个方向

koushiyu commented 1 year ago

快照0001 就是直接通过这种方式,利用非深度学习方法就能获得这些中心是吗,大佬方便详细说说吗,或者指个方向

他的意思是如果在已知中心的前提下,不能从roi出中心啊

dangis-lll commented 1 year ago

那不就还是得沿用第一二阶段提取中心的方法来获得这个中心

koushiyu commented 1 year ago

那不就还是得沿用第一二阶段提取中心的方法来获得这个中心

我理解是这样的

sakthivelj commented 1 year ago

In the semantic segmentation of a single tooth in the third stage, I encountered a problem on my own data and would like to ask for advice. My experiment did not follow the results of the first stage and the second stage, but directly extracted the accurate centroid on the mask and generated the corresponding patch. First of all, the baseline of the data is a single-channel input patch, and the dice, hd, and asd of the test set have basically reached the sota of your experiment; try to concat the skeleton (because considering the propagation error, this skeleton is not even the predicted skeleton, but directly Obtained by gt), the result is that the test set dice has dropped by 0.3 points, hd has remained the same, and asd has dropped slightly; after adding the edge of multi-task, it is the same as the case of concat skl, and there is no obvious improvement. After the experiment, I analyzed that my practice is completely based on the premise that centroid is completely and accurately estimated. I also saw that your previous answer mentioned that in fact, your experimental centroid has added disturbances, and there is no possibility of absolute accuracy. My question is, can the following conclusions be drawn: if centroid can be infinitely close to its true position, then we can also make single-tooth segmentation reach sota without adding skl, edge, and landmark

Hello, have you successfully reproduced this project? In the stage of single tooth segmentation, {'image': image, 'label': label, 'centroid': centroid, 'boundary': boundary, 'keypoints': keypoints} are needed for training. What is boundary and keypoint? How did you get it?

I have same doubt, how to get the boundary, keypoints

koushiyu commented 1 year ago

https://scikit-image.org/docs/stable/api/skimage.segmentation.html#skimage.segmentation.find_boundaries

koushiyu commented 1 year ago

You can even use only the mask for training

sakthivelj commented 11 months ago

Hi, how extract the key points from mask?

DIO-1-2 commented 11 months ago

i have said i don't know, maybe you can ask the author.

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2023年10月1日(星期天) 晚上11:36 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [ErdanC/Tooth-and-alveolar-bone-segmentation-from-CBCT] About the 3rd stage(individual tooth segmentation) (Issue #10)

Hi, how extract the key points from mask?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

sakthivelj commented 11 months ago

i have said i don't know, maybe you can ask the author. ------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2023年10月1日(星期天) 晚上11:36 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [ErdanC/Tooth-and-alveolar-bone-segmentation-from-CBCT] About the 3rd stage(individual tooth segmentation) (Issue #10) Hi, how extract the key points from mask? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

Have you did the stage_1, I have completed I need to verify my result?

sakthivelj commented 6 months ago

Hello, I am having trouble identifying the root tip point and classification the ID of a tooth. If anyone has any information or advice on this topic, please share it with me.

could you please anyone provide additional details? I have been struggling with this issue for a long time. I have only been able to complete the skeleton and boundary of the tooth, but have been unable to identify the root point accurately.

WRRW615 commented 2 weeks ago

第三阶段单颗牙齿的语义分割中,在自己的数据上遇到了一个问题想请教一下。 我这个实验没有沿用第一阶段和二阶段的结果,而是直接在mask上提取了准确的centroid,生成了对应的patch。数据首先baseline是单通道输入patch,测试集的dice、hd、asd就都基本已经达到了您实验的sota;尝试concat了骨架(因为考虑到传播误差,这个骨架甚至不是预测的骨架,而是直接通过gt得到的),结果是测试集dice反而下降了0.3个点,hd持平,asd小幅下降;加入multi-task的edge以后也和concat skl的情况一样,没有明显的提升。 试验后我分析,我这种做法是完全建立在centroid完全准确估计这一前提条件下的。也看到您之前的回答提到了实际上您的实验centroid加入了扰动,不会有绝对准确的可能。 我的问题在于,可否得出如下结论:倘若centroid能够无限接近其真实位置,那我们也能够在不加skl和edge、landmark的情况下使得单牙分割达到sota

你好,请问你在mask上提取准确的质心是通过作者提出的快速聚类实现的吗?得到的质心是一个点吗?