Introduction
In 2021, shortly after the release of the Chinese Mammography Database (CMMD), I conducted my Masters Dissertation research study titled "Mammography Classification Using Deep Learning". This research aimed to explore the potential of applying deep learning techniques to classify mammography images using one of the most recently available and underexplored datasets. My work was among the earliest to implement deep learning models on the CMMD dataset, setting an important foundation for future research in the field.
Motivation
With breast cancer being the leading cause of cancer-related deaths among women, improving diagnostic techniques is crucial. The release of the CMMD dataset presented an exciting opportunity to apply state-of-the-art deep learning approaches to breast cancer screening, a field where innovation can dramatically improve patient outcomes. By engaging with this dataset early on, I sought to contribute to this advancement.
Methodology
I focused on developing a deep learning pipeline for the CMMD dataset, exploring key models such as AlexNet and LeNet, while implementing transfer learning models like ResNet50, VGG16, and Xception. I also performed hyperparameter optimisation to assess the impact of factors like input size and learning rate on classification accuracy.
Results
The deep learning pipeline I developed was one of the earliest applications to the newly released CMMD dataset. The Xception model, along with a fine-tuned version, yielded the best results with 70.61% accuracy and an AUC of 66%. Although these metrics are still below clinical application standards, they serve as early benchmarks for future research efforts.
Conclusion
While my work shows promising initial results, much more needs to be done to reach the accuracy levels required for clinical deployment. However, as one of the earliest research studies on the CMMD dataset, this dissertation sets a benchmark and opens the door for further exploration. All the code is available on GitHub to promote reproducibility and foster collaboration within the research community.
Future Work
Further improvements could be made by using larger datasets, refining models, and optimising hyperparameters. As the CMMD dataset becomes more widely used, I hope to see significant advancements in the use of AI for breast cancer diagnostics.