The process, as performed by us, consists of three steps:

1. Patch extraction (and randomization) (extractPatches.py)

The original MKT (cellVizio) files are read, patches without annotated artifacts are extracted according to their frame entry in the CLE database.

40 batches with randomized patches (and according information about the originating sequence, patient etc.) are generated.

2. Training of a classifier and testing the classifier with the test set (train.py)

TensorFlow is used with a convolutional network, as described in the paper. This step generates a SQLite database with the results of individual frames.

3. Fusion of Patch Probabilities to Image Probabilities (CNN_ppf.py)

In this step, a numpy array with images probabilities is generated.

Unfortunately, to reproduce the exact results of our paper, access to the CLE image database is needed, and we have no permission to provide this openly at this time.

However, to be transparent about the methodology, you can find the code for each of the three steps mentioned above in the following. Further, I provide all derived data, i.e. the results of the classification approach on the patches and the CLE database without the images. With this, it is possible to re-run the patch probability fusion step (step 3). 

We base our CLE image database on a SQLite relational database with the following structure:

   CLEdatabases: Each database consists of a number of sequences (movies),

                 and may be from a different hospital or anatomical region

                 It has the following fields:

                       id:          Unique ID

                       path:        Path to the sequence files (mkt files)

                       description: Description of the database

                       authors:     Medical authors of the dataset

   CLEsequences: Each sequence consists of a number of images (frames),

                 It has the following main fields:

                       id:         Unique ID

                       patientID:  Unique identifier for patient (numerical)

                       patient:    Patient identifier (string)

                       fileId:     Unique identifier for a file (a file may contain in multiple CLE sequences)

                       file:       MKT file name

                       subfolder:  Subfolder containing MKT files

                       database:   Link to CLEdatabases.id

   CLEframes:    Single images of the CLE database.

                 It has the following main fields:

                       id:            Unique ID

                       sequenceID:    Link to CLEsequences.id

                       frameIdx:      Frame index in the MKT file (0: first image in file)

                       cellStructure: Cell classification (-1: unknown, 0: normal

                                      epithel, 1: carcinoma, 2: dysplasia)

                       anatomicalLocation: Location where the sample was taken

                                           (0: upper alveolar ridge, 1: lower inner

                                            labium, 2: palatal region, 3: lesion region,

                                            4: vocal folds)

                       gaussianNoiseClass: Noisyness of the image (0: not, 10: completely)

                       motionArtifactClass: Motion artifacts in image (0: none, 10: only artifacts)

                       illuminationArtifactClass: Illumination artifacts in the image (deprecated)

                       imageQuality:  Subjective quality of image (0: bad, 1: neutral, 2: good)

   CLEregions:   Regions of interest / region annotations within a single frame.

                 Fields:

                       id:            Unique ID

                       frameId:        Link to CLEframes.id

                       regionType:     Type of annotation (0: motion artifact, 1: noise artifact, 2: other artifact,

                                       3: Other ROI/no artifact, e.g. anatomical structures)

                       x1,x2,y1,y2:    Coordinates within image [x1:x2,y1:y2]