What question did this study set out to answer?

To improve breast cancer diagnosis in mammography by integrating complex-moment invariant descriptors with deep learning methods.

April 25, 2026Open Access

Complex-Moment Invariant Descriptors for Enhancing ResNet50-Based Breast Cancer Mammography Classification

Key Points

To improve breast cancer diagnosis in mammography by integrating complex-moment invariant descriptors with deep learning methods.
Developed a hybrid framework combining similarity-invariant descriptors and deep CNN features.
Utilized ImageNet-pretrained ResNet50 for feature extraction without modifying the backbone.
Conducted experiments on CBIS-DDSM with stratified 5-fold cross-validation.
Achieved 94.8% accuracy in mammography classification.
Obtained an AUC of 0.938, indicating strong diagnostic performance.
Enhanced accuracy by +3.6% over the baseline model.

Abstract

Breastcancer screening relies on mammography, yet automated diagnosis remains challenging due to subtle lesion appearance and acquisition variability. While deep CNNs achieve strong performance, geometric invariance is typically learned implicitly. This paper proposes a hybrid framework integrating similarity-invariant descriptors derived from com plex moments with deep convolutional features. The descriptors cancel translation, rotation, and isotropic scaling, and are fused with an ImageNet-pretrained ResNet50 via lightweight late fusion without modifying the backbone. Experiments on CBIS-DDSM using stratified 5-fold cross-validation achieve 94.8% accuracy and 0.938 AUC, a +3.6% gain over the baseline.

Complex-Moment Invariant Descriptors for Enhancing ResNet50-Based Breast Cancer Mammography Classification

Key Points

Abstract

Cite This Study