dpetrini/cbis_ddsm_rev

In this repository we provide an alternative version of CBIS-DDSM Dataset.

BREAKING NEWS (2026-06)

Full source code released of my research: After a while fixing and cleaning the code, I finally released the full source code of my research. You can find it in: https://github.com/dpetrini/multiple-view-src

Introduction

CBIS-DDSM original dataset has 3.103 files of full images of mammography exams. It has patches of lesions but we don´t consider it here.

The original dataset can be downloaded here: https://www.cancerimagingarchive.net/collection/cbis-ddsm/

There is also a paper explaining its history and exams details: https://www.nature.com/articles/sdata2017177

The original dataset is provided in dicom images of resolution ranging from 3.920 to 6.931 (H) x 1.786 to 5.431 (W) pixels. It also provides a test division. We believe a test division is a common ground so that researchers can compare results. CBIS-DDSM has only positive cancer images, it means no normal (not cancer) images. The positive cancer images are of category benign and malign. For full images it has following counting:

Type	Benign	Malign	Total
train	1.347	1.111	2.458
test	381	264	645
Total	1728	1375	3103

In this distribution we provide some enhancements to provide easy to use for machine learning research, like the following features:

• Images converted to PNG with resolution 1.152 (H) x 896 (W).
• Filtering to reduce misunderstandings: some images have two lesions, one benign and other malign, we consider it malign.
• The original dataset is organized in mass and calcification lesions. Some patients have both lesions and eventually some original images are duplicated, like same pacient with lesions in same side, one image for mass and same image for calcification lesions, and they can be in training and test. As we consider dataset as a whole, we erased these duplicated images in test set, leaving them only in training set. When we found same patient but different side (L or R), we moved from test dataset to train dataset ("_rev" files) to keep patient-level consistency.

After points raised above we end up with following counting:

Type	Benign	Malign	Total
train	1.364	1115	2234
test	349	246	595
Total	1713	1361	3074

We suggest a validation partition extracted from training images, so we have:

Type	Benign	Malign	Total
train	1.233	1001	2234
val	131	114	245
test	349	246	595
Total	1713	1361	3074

If you want, join val to train and have the complete train partition.

After training and testing this dataset with single image classification, we produced following results, evaluating in test division:

Network	EfficientNet-B0	MobileNetV4_Small	EfficientNet-B3
AUC	0.8137±0.0179	0.7410±0.0203	0.8212±0.0176

Let us know if you have other results.

Best Regards

D.Petrini

ps. check also our below work, although not using this version of dataset:

Optimizing Breast Cancer Detection in Mammograms: A Comprehensive Study of Transfer Learning, Resolution Reduction, and Multi-View Classification (2025) PETRINI, D. G. P.; KIM, H. Y. Optimizing Breast Cancer Detection in Mammograms: A Comprehensive Study of Transfer Learning, Resolution Reduction, and Multi-View Classification. 2025. arXiv:2503.19945 [eess.IV]. Disponível em: https://arxiv.org/abs/2503.19945.