Hello dear researchers: I am working on project to classify lung ct image using CNN and LIDC dataset, i have three questions:
1- each scan in the dataset contains a large number of slices(images) between 100 to 500, do i need all the image for each scan or i can choose some of them for example 20 images for each scan?
2- what is the best segmentation algorithm for these type of images?
3- each image in the dataset has a size 512x512, to i need to resize images or use them as they are?
Dear Hunar A. Ahmed ,
Look the link, please.
https://beta.vu.nl/nl/Images/werkstuk-heeneman_tcm235-876475.pdf
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5569872/
https://arxiv.org/ftp/arxiv/papers/1611/1611.06651.pdf
Regards, Shafagat
Dear Ahmed,
It all depends on what you want to classify. For instance, if the exam has cancer or not, or if there is lesions or not, or if the lesions are benign or malignant, and so on.
Regards,
Raniery
Dear Hunar A. Ahmed
In general, the input data selection ( in terms of number and dimensions ) depends on the desired network structure , the classification object and related features as well as computational time and cost.
However, I suggest that, in order to evaluate the convergence and accuracy of the network, it is better to train the network with fewer input images. After that, the number of images can be increased step by step.
Given that usually part of each image is considered as a background and does not have much effect on the classification, therefore, by utilizing an appropriate feature extraction method as a pre-processing method(Histogram features, wavelet based methods, ...), the dimensions of the input data can be reduced as well.
best regards
It seems that others have provided a set of good resources to start your research. I also have some comments:
With regards to number of images/slices for each scan, for a 2D approach choose one or two slices per scan that contains large enough area of the suspicious/nodule region. for a 3D approach choose all the slices that contain the suspicious region.
What you do, in terms of choosing segmentation algorithm or image size, really depends on what type of classification you are interested. As mentioned by others, based on your task you decide to choose which network to use ( agree that U-net can be useful, also check for Resnet) then based on your network architecture (the required input dimensions by network) modify the image size.
One think you can keep in mind is that if you want to use CNN for a classification task (e.g. nodule vs. non-nodule) you do not necessarily have to have segmentation and a bounding box around the suspicious region could be enough.
The CNN is best CT image classification. But one thing it takes time consumption. In the case of medical image classification, time consumption is not a matter, but at the same time accuracy has to be maintained. You have to concerned about what technique you are using to train the CNN that is the most important, because the mathematical/statistical techniques determine the precision/accuracy.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images