Dicom Image De-Identifier

This is a simple DICOM image de-identifier. Its primary function is to receive a DICOM file, extract its payload (one or more images), apply image detection to each corresponding image and then remove the corresponding detected areas from that image.

Installation

Execute in a terminal

python3 -m pip install -r requirements.txt

Results

Resulting text detection and removal of text from respective bounding boxes based on the CRAFT detector of Keras OCR.

Available Image Detector/OCR Engines

• Keras OCR
• Tesseract OCR
• Presidio

Text Removal

Pipeline

To select an image text removal pipeline, open main.py, in the line

PIPELINE = <PIPELINE_FUNCTION>

replace <PIPELINE_FUNCTION> with one of the functions within the lines

## ! Pipelines: Begin

...

## ! Pipelines: End

that can be found inside dcm_img_text_remover.py.

For one file conversion you can use

presidio_dicom_image_text_remover
pytesseract_dicom_image_text_remover
keras_ocr_dicom_image_text_remover

For multiple file conversions you can use

MassConversion

Input Files

For One Input File

To select one input DICOM file (with name e.g. pos2.dcm), first place it inside ../dataset/raw and specify its path through the parameter IN_PATH at main.py, e.g.

IN_PATH = 'pos2.dcm'

For Multiple Input Files

For multiple DICOM conversions simply paste your directory path (e.g. ../dataset/raw/direc) and specify its path through the parameter IN_PATH at main.py by placing this line at the beginning of the pipeline's function inside dcm_img_text_remover.py, e.g.

IN_PATH = '../dataset/raw/direc'

Output Files

For One Input File

You can find the cleaned DICOM file along with its prediction plot on the path ./dataset/clean with the corresponding filename as its input. If the plot is unwanted, it can be disabled by commenting out the lines

vis_obj = visuals.DetectionVisuals(...)

vis_obj.build_plt(...)

rw_obj.store_fig(...)

from the associated pipeline function inside dcm_img_text_remover.py.

Run Script

To execute navigate inside ./src and apply

python3 main.py -p <input_directory_path> --gpu

To avoid using GPU, one may remove the --gpu argument.

Technical Description

Pipelines

Keras OCR

A high abstraction of the dicom image text removal pipeline based on keras-ocr. In this demonstrative example, the DICOM file contains exactly one image defined as $I$. The output image $I'$ is the cleaned image, based on the trained CRAFT model's estimation of bounding box locations in the image.