Paper 9
Li, H., Chen, D., Nailon, W. H., Davies, M. E., & Laurenson, D. (2022). Dual Convolutional Neural Networks for Breast Mass Segmentation and Diagnosis in Mammography. IEEE Transactions on Medical Imaging, 41(1), 3–13. https://doi.org/10.1109/tmi.2021.3102622
Li et al. (2022) introduce an innovative dual-path Convolutional Neural Network (CNN) architecture named DUALCORENET for mammogram analysis. Their research leverages the Digital Database for Screening Mammography (DDSM) and INbreast datasets, enhanced through augmentation techniques such as horizontal and vertical flips and random crop. DUALCORENET comprises two distinct paths: the Locality Preserving Learner (LPL) focuses on extracting hierarcThrough an extensive review of 12 research papers, we examine various CNN architectures and methodologies. Key findings include the effectiveness of hybrid models, advanced pre-processing techniques, and the potential of transformer-based approaches.hical and local intrinsic features from large-scale Regions of Interest (ROIs), encompassing textural and contextual information crucial for mass classification, while the Conditional Graph Learner (CGL) specializes in extracting segmentation-related and geometrical features from resulting binary masks. A novel breast mass segmentation CNN architecture, featuring a Conditional Random Field (CRF) inference layer, is integrated to enable precise segmentation at high resolutions. The model optimizes training using the Adam optimizer, and dropout layers are employed for enhanced generalization. The study demonstrates remarkable segmentation and classification performance, achieving Dice Coefficients of 93.69% (INbreast) and 92.17% (DDSM). The study’s strength lies in its pioneering dual-path approach, providing both high-resolution segmentation and accurate classification, offering valuable insights into effective strategies for mammography analysis, thereby enhancing the potential of my own project in this domain.
Paper 10
Jayesh George Melekoodappattu, Anto Sahaya Dhas, Binil Kumar Kandathil, & Adarsh, K. S. (2022). Breast cancer detection in mammogram: combining modified CNN and texture feature based approach. Journal of Ambient Intelligence and Humanized Computing, 14(9), 11397–11406. https://doi.org/10.1007/s12652-022-03713-3
Jayesh George Melekoodappattu et al. (2022) present an innovative method that combines Convolutional Neural Network (CNN) and texture analysis for precise mammogram classification. Their research draws upon the Mammographic Image Analysis Society (MIAS) and Digital Database for Screening Mammography (DDSM) datasets. The methodology begins with pre-processing, involving denoising and enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE). Subsequently, a nine-layer modified CNN is employed for primary mammogram image classification. In parallel, multi-scale local binary patterns (LBP) extract texture features from various window sizes, while Gabor filters identify multi-scale and multi-orientation textural micro-patterns. Linear Discriminant Analysis (LDA) and Uniform Manifold Approximation and Projection (UMAP) are utilized to reduce feature dimensions. The study integrates results from CNN classification and the feature extraction process using an ensemble method for final decision-making. Remarkable results are achieved, with a specificity of 97.8% and accuracy of 98% on MIAS, and 98.3% specificity and 97.9% accuracy on DDSM. The study’s strength lies in its comprehensive approach, combining deep learning and texture analysis to enhance measurement metrics. This method represents a significant step toward more accurate breast cancer detection in mammograms, offering valuable insights into the integration of texture features with CNN for improved mammogram classification.
Paper 11
Elkorany, A. S., & Elsharkawy, Z. F. (2023). Efficient breast cancer mammograms diagnosis using three deep neural networks and term variance. Scientific Reports, 13(1). https://doi.org/10.1038/s41598-023-29875-4
Elkorany and Elsharkawy (2023) present a hybrid technique for breast cancer classification that optimizes efficiency and performance by focusing on patches of interest (ROIs) from the Mammographic Image Analysis Society (MIAS) database rather than processing entire images. This targeted approach enhances feature extraction and classification. The study employs three CNN models (Inception-V3, ResNet50, AlexNet) for feature extraction, with each model offering distinct advantages. Inception-V3 efficiently captures information at various scales using inception modules, ResNet50’s residual connections mitigate the vanishing gradient problem, and AlexNet’s pioneering architecture contributes to its image recognition capabilities. Term Variance (TV) feature selection algorithm is applied to refine and select the most informative features extracted by deep learning models based on their variance across classes. The final classification is carried out by a multiclass support vector machine (MSVM), well-suited for complex classification tasks with high-dimensional data. The study achieves an impressive classification accuracy of 97.45% on the MIAS dataset, demonstrating the effectiveness of the combined deep learning and feature selection approach in accurately classifying breast cancer in mammograms. This approach holds the potential to contribute to the development of efficient and accurate mammogram classification models for my project.
Paper 12
Betancourt, A. S., Marrocco, C., Molinara, M., Tortorella, F., & Bria, A. (2023). Transformer-based mass detection in digital mammograms. Journal of Ambient Intelligence and Humanized Computing, 14(3), 2723–2737. https://doi.org/10.1007/s12652-023-04517-9
Betancourt et al. (2023) tackle the challenge of mass detection in mammograms with a transformer-based approach, leveraging the OMI-DB mammography image database. The study includes crucial pre-processing steps such as image resizing, normalization, and data augmentation to enhance model robustness. The backbone feature extractor chosen is the Swin Transformer, known for its ability to handle various scales and generate high-resolution feature maps. Object detection methods RepPoints and Deformable DETR are employed to further enhance accuracy. RepPoints identifies representative points for object localization, while Deformable DETR, a transformer-based framework, uses deformable attention to focus on key sampling points. This combination aims to improve mass detection accuracy. The study also applies Weighted Boxes Fusion (WBF) for prediction merging, effectively combining overlapping bounding boxes from multiple detectors based on their confidence scores. This results in more accurate and robust object detection, especially beneficial for complex scenarios like mass detection in mammograms. The study achieves a True Positive Rate (TPR) of 75.7% at 0.1 False Positives per Image (FPpI), representing notable improvements over convolutional methods and state-of-the-art approaches. Additionally, combining transformer- and convolution-based detectors with WBF leads to a further TPR improvement, reaching 78.1% TPR at 0.1 FPpI. This research offers valuable insights into alternative feature extraction methods and improved mass detection in mammograms, potentially benefiting my own project in this domain.
Reflection
Reflecting on the 12 papers, there is a noticeable trend towards the integration of advanced CNN architectures with diverse methodologies for breast cancer detection in mammograms. These studies make effective use of various datasets, with a primary focus on the MIAS and DDSM datasets, ensuring a comprehensive analysis of breast cancer detection. Data pre-processing emerges as a critical step in mammogram analysis using CNNs, encompassing tasks such as image enhancement, noise reduction, and data normalization. These pre-processing techniques, including resizing, denoising, and augmentation, play a pivotal role in improving model accuracy and robustness. They enable CNN models to focus on relevant features and handle the real-world variations present in mammogram images, making them essential for the development of reliable and efficient diagnostic tools. The studies showcase innovative approaches such as wavelet decomposition, hybrid CNN models, attention mechanisms, and transformer-based techniques, all of which contribute to achieving consistently high levels of accuracy and specificity. The strengths of these studies lie in their creative combination of techniques and the attainment of high-performance metrics. However, they also acknowledge limitations, including potential dataset biases and the need for broader validation. Overall, these studies offer invaluable insights that can significantly benefit my own project, particularly in the design of effective CNN models and the incorporation of novel techniques to enhance the accuracy of breast cancer detection in mammograms.
