CNN1D-LSTM with SMOTE for breast cancer classification: performance and statistical insights

Area of Science:

• Oncology
• Biomedical Engineering
• Machine Learning

Background:

• Breast cancer diagnosis is critical for effective treatment and survival.
• Early and accurate detection methods are essential for improving patient outcomes.
• Imbalanced datasets pose a challenge in developing reliable diagnostic models.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for automated breast cancer detection.
• To address data imbalance issues in breast cancer datasets using over-sampling techniques.
• To assess the performance of the proposed model across different dataset scenarios.

Main Methods:

• Utilized Convolutional Neural Network 1D (CNN1D) integrated with Long Short-Term Memory (LSTM) for feature extraction.
• Employed the Synthetic Minority Over-sampling Technique (SMOTE) to balance imbalanced datasets.
• Evaluated the model using Support Vector Classification (SVC) and Random Forest Classifiers (RFC) with K-fold cross-validation.

Main Results:

• The CNN1D-LSTM-SMOTE model achieved high diagnostic accuracy, with Matthews Correlation Coefficient (MCC) scores of 97.2% (Dataset-1) and 100.0% (Dataset-2) using SVC.
• Random Forest Classifiers demonstrated superior performance on combined features, achieving an MCC of 98.4%.
• K-fold cross-validation yielded average MCCs of 91.7%, 74.1%, and 96.9% for Dataset-1, Dataset-2, and combined features, respectively, with statistically significant results (p=0.01).

Conclusions:

• The proposed CNN1D-LSTM-SMOTE model demonstrates significant potential for early, reliable, and automated breast cancer diagnosis.
• The integration of deep learning and data balancing techniques offers a promising approach to enhance diagnostic accuracy.
• This model could support oncologists, leading to improved diagnostic procedures and better patient outcomes.