search

nyukat / mammography_metarepository

Meta-repository of screening mammography classifiers
https://arxiv.org/abs/2108.04800
BSD 2-Clause "Simplified" License
65 stars 11 forks source link

Prelim. results, Image Size and contrast #3

Closed jamesjjcondon closed 3 years ago

jamesjjcondon commented 3 years ago

This is only 38 exams / patients. Prior is roughly 40% malignancy (patient-wise):

Evaluating. Image-level metrics Average precision: 0.5525447038757333 AUROC: 0.7534153005464481 AUPRC: 0.547119407370931 ROC Plot: /data1/NYU_metarepo/predictions/nyu_gmic_test01_predictions_image_level_roc_curve.png PRC Plot: /data1/NYU_metarepo/predictions/nyu_gmic_test01_predictions_image_level_pr_curve.png Breast-level metrics Average precision: 0.6278010396248606 AUROC: 0.7551912568306012 AUPRC: 0.6212351976579777 ROC Plot: /data1/NYU_metarepo/predictions/nyu_gmic_test01_predictions_breast_level_roc_curve.png PRC Plot: /data1/NYU_metarepo/predictions/nyu_gmic_test01_predictions_breast_level_pr_curve.png Prediction file: /data1/NYU_metarepo/predictions/nyu_gmic_test01_predictions.csv

Nice work!!!!

Check this out:

Original image size: NYU-left-BSSA-right

Contrast differences: BSSA_left-NYU-right

I'm about to publish re. I'll run on a larger set, hopefully tomorrow, but I think the above is enough to shift the statistics of the input data and therefore the result / transferability, hence my suggestion about a "pre-model" to deal with this, if possible.

Thanks again. Talk soon.

jamesjjcondon commented 3 years ago

Ps - I was hoping to know the original dimensions of your dicom.pixel_arrays which was not in Wu 2019/2020, from memory. So that this can be matched as close as possible.

jwitos commented 3 years ago

@jamesjjcondon Original image dimensions were various, mostly in the 2000-3000 px range, e.g. 3518x2800. I don't have the exact statistics on what were the most common dimensions unfortunately. From the paper I can tell you that CC views were cropped to 2677x1942 pixels and MLO views to 2974x1748 pixels. Also, you might find more interesting details about image cropping etc. in our data report that gives plenty of details on dataset preprocessing. Looking forward to reading the results of your work!