ISIC Forum

Task 3 supplemental information


(Brian Helba) #1

Greetings Task 3 participants,

We are providing some supplemental information, which you may find helpful when splitting the Part 3 Training data for your own internal training / evaluation processes. Use of this data is optional, and no such data will be provided for the formal Part 3 Validation and Test phases, where predictions have to be made based on single image data only.

This supplemental information may be downloaded as a CSV file here.

The structure of this supplemental information is as follows:

  • For each image in the Part 3 Training set (labeled in this CSV as the column "image"), there is a lesion identifier (the "lesion_id" column) and a diagnosis confirm type methodology (the "diagnosis_confirm_type" column).

  • Images with the same lesion identifier value show the same primary lesion on a patient, though the images may be taken at different camera positions, lighting conditions, and points in time.

  • Images with a more rigorous diagnosis confirm type methodology are typically more difficult cases for human expert clinicians to evaluate, particularly when the images’ ultimate diagnosis is benign. In ascending order of rigorousness as applied to cases of the present dataset, the diagnosis confirm type methodologies are "single image expert consensus", "serial imaging showing no change", "confocal microscopy with consensus dermoscopy", "histopathology".

    • So, for example, an image with a diagnosis of "NV" ("Melanocytic nevus") which was confirmed by "histopathology" would typically be a more ambiguous or difficult case for human experts than a similar "NV" ("Melanocytic nevus") image confirmed by "single image expert consensus" only, as the former image required a more invasive procedure to diagnose.

Should you choose to utilize this data in your Part 3 algorithm training process, it is your responsibility to figure out how to best incorporate it. Of course, you are free to discuss and collaborate with other participants via the Challenge Forum.


Starter solution based on Keras/Tensorflow
Cant find the leasion group csv
Task 3 lesion overlap between training and validation set
(Brian Helba) #2