Breast Cancer Diagnosis with AI

 

Breast cancer remains one of the most prevalent and life-threatening diseases worldwide, necessitating accurate and reliable diagnostic methods for early detection. This research introduces a novel diagnostic framework for breast cancer detection using 3D tomosynthesis mammography. By integrating radiomic features with deep learning-derived features from an attention-based 3D autoencoder, the study addresses the challenges of diagnostic accuracy, feature reproducibility across imaging scanners, and clinical interpretability. The proposed method represents a significant step forward in computer-aided diagnosis, offering improved performance through feature harmonization, dimensionality reduction, and ensemble learning approaches.

Dataset and Feature Extraction Techniques
The study utilized a comprehensive dataset of 6672 cases classified into normal, benign, and malignant categories. Radiomic features were extracted from the images and harmonized using ComBat to mitigate scanner-based variability. Deep learning features were derived from an attention-based 3D autoencoder, designed to focus on diagnostically relevant regions within the scans. To ensure reliability, feature reproducibility was evaluated using intra-class correlation coefficients (ICC), retaining only features with ICC > 0.85 for subsequent analysis.

Dimensionality Reduction and Feature Selection Strategies
Given the high-dimensional nature of the extracted features, multiple dimensionality reduction and feature selection techniques were employed to refine the feature space. Methods such as Principal Component Analysis (PCA), Least Absolute Shrinkage and Selection Operator (LASSO), and Recursive Feature Elimination (RFE) were applied to remove redundancy and enhance model efficiency. This process played a crucial role in improving computational performance while maintaining diagnostic accuracy.

Machine Learning Models and Ensemble Learning Approach
The refined features were used to train multiple machine learning models, including XGBoost, CatBoost, and Random Forest. Additionally, a stacking ensemble model was developed, combining the strengths of individual classifiers. A stratified 70/30 train-test split and 5-fold cross-validation ensured robust evaluation of the models, with performance assessed using metrics such as AUC, accuracy, and sensitivity.

Performance Evaluation and Model Comparisons
Among the tested models, the stacking model combined with RFE demonstrated the highest performance, achieving an AUC of 95.75%, accuracy of 94.29%, and sensitivity of 94.74%. XGBoost with PCA also achieved strong results, with an AUC of 94.40%. Notably, the attention-based autoencoder significantly improved deep feature quality, outperforming the non-attention version with a testing AUC of 92.32% and accuracy of 91.22%. The integration of radiomic and deep learning features consistently outperformed models using either feature type alone.

Clinical Relevance and Future Research Directions
The proposed diagnostic framework is not only accurate but also interpretable, addressing key concerns in clinical adoption. By ensuring feature reproducibility and focusing on clinically meaningful regions, the approach enhances trust among healthcare professionals. Future research should explore the integration of multimodal data, such as histopathology and genomic profiles, to further improve predictive performance. Additionally, optimizing the framework for real-time clinical deployment could pave the way for its adoption in routine breast cancer screening programs.

Technology Scientists Awards

===================
Website : technologyscientists.com
Nomination Link : https://technologyscientists.com/award-nomination/?ecategory=Awards&rcategory=Awardee
To Contact : support@technologyscientists.com

#BreastCancerDetection
#MedicalImaging
#Radiomics
#DeepLearning
#3DAutoencoder
#AttentionMechanism
#Tomosynthesis
#MachineLearning
#XGBoost
#CatBoost
#RandomForest
#FeatureSelection
#PCA
#LASSO
#RFE
#EnsembleLearning
#ClinicalAI
#MedicalDiagnostics
#CancerResearch
#AIinHealthcare