paulritsche / deepacsa Goto Github PK

HI Paul,

Great job with DeepACSA!

I am trying install and run your software but found an error I suppose is related with the icon of used in the gui:

(base) PS C:\QuantitativeUS\DeepACSA> & C:/ProgramData/Anaconda3/python.exe c:/QuantitativeUS/DeepACSA/DeepACSA_tkinter.TclError: bitmap "icon.ico" not defined (base) PS C:\QuantitativeUS\DeepACSA> conda activate base (base) PS C:\QuantitativeUS\DeepACSA> & C:/ProgramData/Anaconda3/python.exe c:/QuantitativeUS/DeepACSA/DeepACSA/deep_acsa_gui.py Traceback (most recent call last): File "c:\QuantitativeUS\DeepACSA\DeepACSA\deep_acsa_gui.py", line 351, in <module> DeepACSA(root) File "c:\QuantitativeUS\DeepACSA\DeepACSA\deep_acsa_gui.py", line 36, in __init__ root.iconbitmap("icon.ico") File "C:\ProgramData\Anaconda3\lib\tkinter\__init__.py", line 2080, in wm_iconbitmap return self.tk.call('wm', 'iconbitmap', self._w, bitmap)_tkinter.TclError: bitmap "icon.ico" not defined

I also had some issues with some gui_helpers, it seems that the new version of Tensorflow has changed where "img_to_array" lives.

Loading personal and system profiles took 2294ms. (base) PS C:\QuantitativeUS\DeepACSA> conda activate base (base) PS C:\QuantitativeUS\DeepACSA> & C:/ProgramData/Anaconda3/python.exe c:/QuantitativeUS/DeepACSA/DeepACSA/deep_acsa_gui.py Traceback (most recent call last): File "c:\QuantitativeUS\DeepACSA\DeepACSA\deep_acsa_gui.py", line 11, in <module> import gui_helpers File "c:\QuantitativeUS\DeepACSA\DeepACSA\gui_helpers\__init__.py", line 10, in <module> from gui_helpers.predict_muscle_area import * File "c:\QuantitativeUS\DeepACSA\DeepACSA\gui_helpers\predict_muscle_area.py", line 20, in <module> from keras.preprocessing.image import img_to_array ImportError: cannot import name 'img_to_array' from 'keras.preprocessing.image' (C:\ProgramData\Anaconda3\lib\site-packages\keras\preprocessing\image.py)

I managed to sort this "'img_to_array' error out but not sure if my .ico error is because of me and my environment installation or something else that has to be updated in the code.

Sorry if these are fairly naïve questions but I am new at Python.

Cheers,

PPM

In order to evaluate the models predictions, it is of benefit if the original US images are displayed next to the suggested masks.
So far, this is done usign the preprocessed images. If instead the unprocessed original images could be used, it could help the user to better evaluate the predictions.

In order to analyze the GM & GL muscles too, the rigth hardcoded parameters for the Ridge detection (used during automatic scaling) need to be implemented. Therefore, Ridge detection has to be tested on some GM & GL images with different settings.

Add as a try statement when EFOV hugh line transform won't work. This is because most images don't have scaling line.

Implement k-fold crossvalidation to evaluate best set of hyperparameters for model training.

Should allow for breaking of program while running, and to allow reuse of interface without closing.

In order to accurately predict the muscle area, image pre-processing might be of help.
Therefore, histogram equalization, more specifically CLAHE, should be implemented in model training and prediction.
CLAHE should be applied to the imported image before resizing and reshaping.

Additionally to predicting the muscle area, it would be great if echo intensity could be calculated.
Echo intensity classifies the mean pixel intensity (on a grayscale). It might be a useful parameter for describing several physiological muscle characteristics. However, the evidence is not quite clear about its applicability yet. Nonetheless, it is an inmprovement to the code.
The echo intensity should be calculated within the predicted area.

When trying to activate the Deep_ACSAuto environemnt in conda, the command

conda activate Deep_ACSAuto

does not work on all environments. If this happens, the following alternative should be used:

source activate Deep_ACSAuto

Given that images of different devices have different scaling bar features, this should be regarded in the scaling process. Thus, the "bar" option needs to be adapted.

Integrate image cropping function used in model training into prediction script.
Create new function (from jupyter notebook) and file in python and use it after loading the image (crops unnecessary US-Infos). Might be less efficient, but one step less for the user.
Images the model is trained on and images used for prediction should be pre-processed in the same way