Data Collection
This project will leverage a total of 2005 high-resolution open-source mammograms, comprising 1003 contrast-enhanced spectral mammograms (CESM) and 1002 subtracted mammograms from Khaled et al.’s dataset (2021). Within the 1003 CESM, there are 341 normal, 331 benign, and 331 malignant mammograms. Simultaneously, the 1002 subtracted mammograms include 415 normal, 256 benign, and 331 malignant mammograms. As described by Khaled et al. (2021), CESM employs standard digital mammography equipment with specialized software for dual-energy image acquisition. Post-intravenous injection of the patient with non-ionic low-osmolar iodinated contrast material (at a dose of 1.5 mL/kg), craniocaudal (CC) and mediolateral oblique (MLO) views are captured two minutes later. Each view encompasses two exposures, one with low energy (peak kilo-voltage values ranging from 26 to 31 kVp) and the other with high energy (45 to 49 kVp). Subsequent to image acquisition, advanced image processing techniques are applied for recombination and subtraction of low and high-energy images, effectively suppressing the background breast parenchyma to generate subtracted mammograms. All mammograms have been converted from DICOM to high-resolution JPEG images with an average size of 2355 x 1315 pixels.
Data Pre-Processing
In the initial phase, a meticulous screening process will be applied to all mammograms, eliminating images with artifacts from the dataset. Following this, the mammograms will be extracted and categorized into normal, benign, and malignant classes. Subsequently, to ensure uniformity, each mammogram will be converted to grayscale 8 bits/channel JPEG images and resized to a standardized dimension by adding a 1500 x 2250 black colour canvas as a background, effectively preventing any distortion of the mammogram resolution. The pre-processing stage will then proceed with data augmentation, involving flipping each mammogram upside down, left to right, or both, thereby expanding the dataset to enhance the model’s generalization and robustness. Within the dataset, 20% of the mammograms are allocated for validation, 70% are utilized for training purposes, while the remaining 10% are used as test set.
Deep Learning Technique
The pre-processed mammogram undergoes analysis within Python, leveraging deep learning libraries, including TensorFlow with Keras, for the construction of the Convolutional Neural Network (CNN). The model initiates with a series of convolutional layers designed to capture intricate patterns and features within the mammogram images. Following the convolutional layers, pooling layers are introduced to diminish spatial dimensions and extract key features. Subsequently, the flattened output is linked to fully connected layers, functioning as a classifier to discern whether the breast exhibits normal, benign, or malignant characteristics. During training, batches of mammogram images are presented, and the model learns to map inputs to classes through an iterative process, optimizing with the Adam optimizer and ‘sparse_categorical_crossentropy’ loss function. Multiple training epochs refine the model parameters, and validation data ensures generalization.
Model Evaluation
The performance of the Convolutional Neural Network (CNN) model is assessed through various metrics:
- Accuracy: Measures the overall correctness of the model by calculating the ratio of correctly predicted instances to the total number of instances.
- Precision: Assesses the accuracy of positive predictions made by the model, representing the ratio of correctly predicted positive instances to the total predicted positives.
- Recall: Evaluates the ability of the model to capture all relevant positive instances, expressed as the ratio of correctly predicted positive instances to the total actual positives.
- F1 Score: Provides a balanced measure, taking into account both precision and recall through the harmonic mean. It considers false positives and false negatives.
- Receiver Operating Characteristic (ROC) Curve: Offers a graphical representation of the model’s ability to discriminate between classes across various thresholds, plotting the true positive rate (sensitivity) against the false positive rate.
- Cross-Validation: Utilizes k-fold cross-validation to assess the model’s generalization ability and performance across different subsets of the dataset, providing more robust and reliable evaluation metrics.
Hardware Tool
The study makes use of a MacBook Pro equipped with the powerful Apple M1 Max chip, boasting 10 cores intelligently divided into 8 for performance and 2 for efficiency, coupled with a substantial 64 GB of memory. This advanced hardware configuration plays a pivotal role in facilitating the smooth execution of tasks, ensuring optimal performance throughout the study.
Software Tool
For data pre-processing tasks, Adobe Photoshop 2023 was utilized. This version of Adobe Photoshop provided powerful tools for image manipulation and enhancement, ensuring the quality and suitability of the dataset for the CNN training process. Meanwhile, for the construction of Convolutional Neural Networks (CNN), Python version 3.12.0 was employed as the programming language. The CNN implementation leveraged the capabilities of Python, a versatile and widely-used language in the field of machine learning.
Ethical Issues
The project is currently utilizing the dataset provided by Khaled et al. (2021). No additional surveys, interviews, or other forms of user or organizational information will be collected. Consequently, no specific ethical considerations are necessary for data collection or interactions with subjects. To uphold intellectual property, responsible publication practices, and academic integrity, the project will adhere to accurate analysis and reporting, ensuring all sources are appropriately cited and referenced according to the APA7 style.
