Feature extraction from 2D(Ultrasonic, mammogram and MRI image) without annotation /ground truth from radilogists.

[Epimack Michael]

Good day developer and Pyradiomics Communities

I have 300 ultrasonic images, 300 mammogram images and next month I will get 300 MRI images, all of these images are breast images, with the intention of classifying /detection of cancer.

However, these images do not contain annotation (ground truth/mask) from radiologists, but it has the description which images have cancer and which do not, all images are JPG format. I have the following questions.

1. Does it make sense to extract features using pyradiomics, without having annotation from a doctor/ radiologists, based on automatic segmentation to get images mask.

2. There is an open source code that I can convert jpg to NRRD format?

3. Have no great experience on medical images especially doing analysis, can I get open source code for doing analysis and how many images is required to have a good performance?

[Joost van Griethuysen]

Hello,

Ad 1) Depends on how good your auto-segmentation tool is I suppose. From the technical side there is no objection. However, the results will reflect the region you segmented and if that is incorrect, your results may lead to poorer performance. Only way to discover this is to try I think.

Ad 2) .jpg images are also readable by SimpleITK, though it does require some code to select a single color channel (PyRadiomics does not support color images). That said, be aware that jpg images do not have the same geometric information that is present in DICOM images. So if you can get access to the DICOM source images, I'd very heavily advise you use those to convert. Moreover, current conversion tools that work with DICOM can sort the slices stored in separate files. For .jpg source, you'd need to sort them manually, then combine.

Ad 3) Both in R and in Python there are very good packages to do Radiomics analysis with. My experience is mainly in Python, where I generally use packages such as pandas (data handling/sorting), numpy (math functions), scipy (statistical functions), scikit-image (image processing functions) and scikit-learn (machine learning).

Regards,

Joost

Op maandag 30 september 2019 05:31:24 UTC+2 schreef Epimack Michael:

[Epimack Michael]

Thank you for your advise, I will try my best to have DICOM images.