AI-Powered Breast Cancer Prediction: A Deep Learning Approach to Early Detection

Introduction

Breast cancer is one of the most common types of cancer affecting women worldwide. Early detection and diagnosis are crucial for successful treatment and improved outcomes. Traditional methods of breast cancer diagnosis, such as mammograms and biopsies, can be invasive and time-consuming. In recent years, Artificial Intelligence (AI) has emerged as a promising solution to enhance the early detection and prediction of breast cancer by improving diagnostic accuracy, reducing human error, and allowing for personalized treatment plans. This case study explores the application of AI in breast cancer prediction, with a focus on deep learning models, machine learning techniques, and their integration into clinical workflows.

Problem Statement

Now-a-days doctors and hospitals are facing problems and struggling to detect breast cancer earlier to save human life. But the traditional process is time consuming for the serious patient and costly for the economically average person. To remove those problems we develop a predictive model capable of accurately classifying breast cancer patients into benign, malignant, or normal categories based on various clinical and pathological features. This model would aid healthcare professionals in making informed decisions regarding diagnosis, treatment planning, and prognosis.

The Solution: AI Powered Breast Cancer Prediction

AI algorithms, particularly machine learning and deep learning techniques, have the potential to revolutionize breast cancer diagnosis. By analyzing vast datasets of medical images and patient information, AI models can learn to identify patterns and features indicative of breast cancer.

Used key component of an AI-powered breast cancer prediction system: 

• Data Acquisition and Processing:
• Collection of a large MRI Mammography image dataset. We collected 1578 MRI Mammography images including Mask images. Where Benign have 891 images, Malignant have 421 images, Normal have 266 classes dataset available.
• Dropping the corrupted images, Normalization and Augmentation of the images are used for improving the model performance.
• Feature Extraction:

• Extracting relevant features from mammograms, such as converting the same shape of each image into 256 width and height.
• For minority classes, used custom data transformations “Random over-sampling” for reducing balance, imbalance problems during training.
• Dataset Divide:

• Divided the dataset for Training, Validation, and Testing. Where 80% data for Training, 10% data for Validation, 10% data for Testing.
• Model Architecture:

• Used Pretrained ResNet-101 model, which is the convolutional neural network is 101 layers deep.
• Model Training and Validation:

• The ResNet-101 model is fine tuned with the Mammography Breast Cancer dataset.
• Used 20 epochs for model training. After 5 epochs Early Stopping was activated and stopped training.
• After training the model, after 5 epochs Training-Accuracy=0.9160, Training-Loss=23.81% and Validation-Accuracy=0.8169, Validation-Loss=0.3929
• Validation dataset is used to validate the trained model. For Benign, Malignant, Normal f1-scores are 0.88, 0.81, 0.80 respectively.
• Prediction and Interpretation:

• Tested the test dataset with the trained downloaded AI model where benign, malignant, normal accuracy are 0.883721, 0.864865 and 0.787879 accordingly.

Benefits of AI-powered Breast Cancer Prediction

The proposed prediction model offers several potential benefits:

• Early Detection: Accurate prediction can enable early detection of breast cancer, leading to timely intervention and improved treatment outcomes.
• Improved Diagnosis: The model can assist healthcare professionals in making more informed diagnostic decisions, reducing the risk of misdiagnosis.
• Personalized Treatment: By identifying specific cancer subtypes, the model can contribute to personalized treatment plans, tailoring therapy to individual patient needs.
• Resource Optimization: Accurate prediction can help optimize healthcare resources by focusing on patients with the highest risk of developing breast cancer.

The Challenges and Limitations: 

Despite advancements in medical technology, breast cancer diagnosis remains a complex task. False positives and false negatives can lead to unnecessary procedures and missed opportunities for early intervention, respectively. Moreover, the interpretation of medical images, such as mammograms, requires expertise and can be subjective. Breast cancer diagnosis relies heavily on imaging techniques, such as mammograms and histopathological images, and manual interpretation by radiologists or pathologists. This process can be time-consuming, prone to human error, and often leads to missed diagnoses or unnecessary interventions. The key challenges in breast cancer prediction include:

• Data Privacy: Access to large, high-quality medical datasets is crucial for training AI models. Concern around data privacy and patient confidentiality can limit the availability of such datasets.
• Data Quality: The quality and quantity of the available data can significantly impact the model's performance. Incomplete or biased data can lead to inaccurate predictions.
• Data Limitation: AI models always want a huge number of datasets/images. In AI medical applications, getting huge data is one of the toughest points for developing. As our dataset has only 1578 images, that is a really very small amount of images. Need to increase this number.
• Model Complexity: Complex models may require extensive training and computational resources, limiting their scalability and usability in real-world clinical settings.
• Generalization: AI models may perform well on specific datasets but fail to generalize to different populations or imaging devices. Ensuring that models are robust across diverse datasets is crucial for real-world adoption.
• Regulation: The regulatory framework for AI in healthcare is still evolving. Approval from medical regulatory bodies is necessary before AI-based tools can be widely deployed in clinical settings.
• Ethical Considerations: The use of predictive models in healthcare raises ethical concerns regarding data privacy, bias, and the potential for discrimination.

Technologies Used:  

• Deep Learning Framework: We used the most popular research base deep learning framework PyTorch, for building the model and also tested different pretrained model.
• Image Prepocessing libraries: For image preprocessing we used PIL, OpenCV, Matplotlib, Seaborn, Numpy, Scikit-learn for dataset dividation.
• ML Models: We tested different models like Resnet, MobileNet, XCeption and Model built from scratch.
• CUDA-enabled GPU: For efficient processing of AI models we used T4 X2 GPU for training, testing and validation.
• Cloud Computing Platform: HuggingFace Inference Endpoint is used for deployment.

Conclusion:

AI has the potential to significantly impact the field of breast cancer diagnosis. By leveraging the power of machine learning and deep learning, AI-powered systems can enhance the accuracy, efficiency, and accessibility of breast cancer screening. As AI technology continues to advance, we can expect to see even greater improvements in the early detection and treatment of this disease.

https://www.who.int/news-room/fact-sheets/detail/breast-cancer