[Paper Review] Mammograms Classification: A Review

Marawan Elbatel|arXiv (Cornell University)|Mar 4, 2022

AI in cancer detection4 citations

TL;DR

This review evaluates deep learning and classical machine learning approaches for classifying malignant and benign breast masses in mammograms, comparing manually segmented and automatically segmented regions of interest. It highlights that deep learning models, particularly ensemble and weakly supervised networks, have surpassed radiologists in accuracy (AUC +0.11) and outperform traditional methods.

ABSTRACT

An advanced reliable low-cost form of screening method, Digital mammography has been used as an effective imaging method for breast cancer detection. With an increased focus on technologies to aid healthcare, Mammogram images have been utilized in developing computer-aided diagnosis systems that will potentially help in clinical diagnosis. Researchers have proved that artificial intelligence with its emerging technologies can be used in the early detection of the disease and improve radiologists' performance in assessing breast cancer. In this paper, we review the methods developed for mammogram mass classification in two categories. The first one is classifying manually provided cropped region of interests (ROI) as either malignant or benign, and the second one is the classification of automatically segmented ROIs as either malignant or benign. We also provide an overview of datasets and evaluation metrics used in the classification task. Finally, we compare and discuss the deep learning approach to classical image processing and learning approach in this domain.

Motivation & Objective

• To analyze and compare classical image processing and deep learning techniques for mammogram mass classification.
• To evaluate the performance of computer-aided diagnosis (CAD) systems in classifying benign and malignant masses using manually provided and automatically segmented ROIs.
• To assess the impact of dataset choice, evaluation metrics, and model robustness in mammogram classification research.
• To investigate the challenges of explainability, generalization, and computational cost in deep learning-based CAD systems.
• To explore future directions such as federated learning for privacy-preserving, generalized model training.

Proposed method

• The review categorizes mammogram classification into two types: (1) classification of manually segmented ROIs and (2) automated segmentation and classification of ROIs from whole mammograms.
• Classical machine learning methods include feature extraction (e.g., texture, shape, intensity) followed by classifiers like SVM or random forests.
• Deep learning models such as DenseNet-121, ResNet-22, and ensemble architectures (e.g., DMV-CNN) are evaluated for their performance in detecting malignant vs. benign masses.
• Weakly supervised learning is used in models like GMIC, where saliency maps are generated from global and local networks using image-level labels without pixel-level annotations.
• Heatmaps from patch-level classifiers are fused into input channels to improve localization and classification accuracy.
• A gated attention mechanism is applied to fuse global and local feature representations for final prediction.

Experimental results

Research questions

• RQ1 How do deep learning models compare to classical machine learning methods in classifying mammogram masses?
• RQ2 What is the performance difference between manually segmented ROIs and fully automated ROI segmentation in mass classification?
• RQ3 Can weakly supervised deep learning models achieve radiologist-level performance without pixel-level annotations?
• RQ4 What are the key challenges in deploying deep learning models in clinical settings, particularly regarding explainability and generalization?
• RQ5 How can federated learning improve model robustness across diverse imaging devices and populations?

Key findings

• Deep learning models, particularly ensemble and weakly supervised architectures, have achieved higher AUC than radiologists, with one model surpassing human performance by 0.11 AUC.
• The DMV-CNN model, which fuses representations from multiple mammographic views (CC and MLO), improves classification by averaging predictions across views.
• Incorporating heatmaps from a patch-level classifier into the input of a deep network enhanced localization and improved classification accuracy.
• The GMIC model achieved high performance using only image-level labels through a greedy aggregator function to backpropagate gradients from saliency maps.
• Despite high performance, deep learning models suffer from poor interpretability, high computational cost, and limited generalization across different imaging devices and populations.
• Federated learning is emerging as a promising approach to train robust, privacy-preserving models across decentralized, heterogeneous clinical datasets.