AI-DRIVEN BREAST CANCER DIAGNOSIS AND TREATMENT OPTIMIZATION: A DEEP LEARNING FRAMEWORK ALIGNED WITH PAKISTAN'S NATIONAL ACTION FRAMEWORK FOR NON-COMMUNICABLE DISEASES 2021–2030

Abstract

Breast cancer represents a significant public health challenge in Pakistan, where delayed detection, inadequate screening infrastructure, diagnostic disparities, and limited access to personalized oncology services contribute to unfavorable clinical outcomes. This study develops an advanced artificial intelligence (AI)-driven framework for breast cancer diagnosis, disease-stage classification, and treatment optimization, aligned with Pakistan's National Action Framework for Non-Communicable Diseases 2021–2030. The proposed framework integrates Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, attention mechanisms, and Explainable Artificial Intelligence (XAI) to facilitate multimodal diagnostic assessment and personalized treatment decision support. A multicenter dataset comprising 18,500 breast cancer examinations from 5,200 patients was utilized, incorporating mammographic images, breast ultrasound, histopathological imaging, demographic information, clinical biomarkers, and treatment records. Comprehensive preprocessing involved image normalization, lesion segmentation, data augmentation, feature standardization, and patient-level data partitioning into training (70%), validation (15%), and testing (15%) subsets. Experimental findings demonstrated that the proposed attention-enhanced CNN–BiLSTM model achieved a diagnostic accuracy of 97.8%, sensitivity of 98.2%, specificity of 96.9%, precision of 97.4%, F1-score of 97.8%, and receiver operating characteristic area under the curve (ROC–AUC) of 0.989. Comparative evaluation demonstrated improvements over conventional CNN (91.6%), ResNet50 (94.3%), and Random Forest (89.7%) architectures. Furthermore, the integrated treatment optimization module achieved 93.4% accuracy in predicting treatment response categories, while the disease-stage classification component attained 95.1% accuracy. Gradient-weighted Class Activation Mapping (Grad-CAM) enhanced model interpretability by highlighting diagnostically relevant image regions, supporting transparent clinical assessment. Beyond predictive performance, the framework integrates risk-based screening prioritization, referral optimization, and resource-sensitive clinical decision support to address structural challenges in Pakistan's oncology services. The findings highlight the potential of multimodal deep learning to strengthen diagnostic precision, facilitate earlier clinical intervention, support individualized treatment planning, and improve evidence-informed cancer management. This research establishes a computational foundation for integrating explainable AI into breast cancer care while advancing accessible, equitable, and policy-aligned oncology services across Pakistan's healthcare system.