Adapt github project to my nifty files
Budget: €30 – €250 EUR
I need a python software engineer specialist in tensorflow and pytorch to help me. I am working in image segmentation of lesions on MRI images of patients. I have nifty files containing the 3D volume array of the patient´s brain and the corresponding nifty file with its segmentation, same name but different folder. One patient may have several files of different studies in differente time points. I need to adapt the architecture described in this page: https://github.com/robinvvinod/unet to run in google colab environment with my own data.
To avoid memory issues I have developed a patch obtaining add-on that takes nifty files and get patches from them as npy files 32x32x32x1 but datagenerator on the fly from full numpy arrays would be OK too.
First:build a model U-net using 3D convolutions and attention mechanism as an image segmentation pipeline.
Second: use the second time point image/segmentation to “influence/use as starting weight” for attention in order to increase accuracy of the model, i.e having the previous time point image and segmentation use this information to initialize attention function/tensor to improve image segmentation. Pipelines woul be then:
(Timepoint1Image+SegmentationTimePoint1+TimePoint2Image) -> U-Net with attention mechanism -> (TimePoint2Segmentation)
To avoid memory issues I have developed a patch obtaining add-on that takes nifty files and get patches from them as npy files 32x32x32x1 but datagenerator on the fly from full numpy arrays would be OK too.
First:build a model U-net using 3D convolutions and attention mechanism as an image segmentation pipeline.
Second: use the second time point image/segmentation to “influence/use as starting weight” for attention in order to increase accuracy of the model, i.e having the previous time point image and segmentation use this information to initialize attention function/tensor to improve image segmentation. Pipelines woul be then:
(Timepoint1Image+SegmentationTimePoint1+TimePoint2Image) -> U-Net with attention mechanism -> (TimePoint2Segmentation)