Breast cancer remains the most prevalent malignancy among women worldwide and continues to represent a leading cause of cancer-related mortality [1]. Despite substantial advances in screening methods and systemic therapies, surgical resection remains the cornerstone of treatment for localized disease. The principal surgical goal is to achieve clear or negative margins, indicating that no residual malignant cells remain at the excision boundary. Achieving this intraoperatively is critical, as residual tumor presence increases the risk of recurrence and frequently necessitates secondary surgeries, adversely affecting both patient prognosis and quality of life.

Intraoperative approaches include Frozen Section Analysis (FSA) and Imprint Cytology (IC), which allow for immediate, tissue-based pathological feedback. FSA offers histopathological confirmation within approximately 30 min, but it requires specialized pathology personnel and infrastructure within the operating suite. Imprint Cytology, though faster, is labor-intensive, limited by sampling scope, and prone to subjective interpretation [2,3].

Postoperative evaluation, the gold-standard approach, involves detailed histopathological examination of the entire excised specimen. Although highly accurate, this process may take several days, meaning the definitive margin status becomes known only after surgery.

Because of these limitations, surgeons often rely intraoperatively on visual inspection and tactile feedback to estimate margin integrity. Consequently, re-excision rates for breast-conserving surgery (BCS) remain substantial, ranging from 20 % to 30 % worldwide. In Egypt, breast cancer similarly dominates the female cancer landscape and accounts for the highest incidence among women [4]. The time delay and subjectivity associated with existing methods underscore an urgent need for rapid, objective, and label-free margin-assessment technologies that can be integrated directly into surgical workflows [5].

To meet this clinical need, numerous imaging and spectroscopic modalities have been proposed. Hyperspectral imaging (HSI) platforms, for instance, have demonstrated high diagnostic accuracy for automated breast-cancer detection—reporting sensitivity and specificity rates exceeding 95 % through advanced machine-learning algorithms [6]. Photoacoustic microscopy (PAM) has also shown promise as a real-time, label-free modality capable of visualizing microvascular structures with histology-like detail, effectively differentiating malignant from normal tissue intraoperatively [7]. Beyond optical methods, liquid biomarker-based diagnostics have advanced as complementary radiation-free strategies, leveraging circulating nucleic acids, microRNAs, extracellular vesicles, and other metabolic indicators for early detection [8]. Despite these innovations, challenges persist in balancing sensitivity, speed, portability, and ease of intraoperative deployment.

Among optical techniques, diffuse reflectance near-infrared (DR-NIR) spectroscopy has emerged as a particularly powerful candidate for intraoperative applications detection [9]. DR-NIR provides quantitative insight into tissue composition by analyzing wavelength-dependent absorption and scattering, thereby distinguishing variations in hemoglobin, lipid, and water content that differentiate malignant from benign tissues [10]. The integration of micro-electromechanical-systems (MEMS)-based spectrometers, notably the NeoSpectra-Micro sensor [11], has propelled this field further by delivering laboratory-grade spectral performance in a compact, cost-effective, and portable format [12].

In parallel, artificial intelligence (AI) and machine learning (ML) have profoundly enhanced medical-data interpretation, enabling automated extraction of subtle diagnostic patterns from optical spectra [13]. Traditional ML algorithms—such as support-vector machines, logistic regression, and ensemble classifiers—excel at identifying deterministic spectral boundaries, while deep-learning (DL) architectures (e.g., convolutional neural networks, CNNs) can capture nonlinear, hierarchical relationships, offering improved robustness to noise and heterogeneity [14]. However, most prior work either evaluates sensor performance in isolation or focuses on algorithmic benchmarking divorced from real-time clinical integration.

To bridge these gaps, the current study presents a clinically oriented, fully integrated DR-NIR + AI framework for rapid intraoperative breast-cancer margin assessment. The system combines a MEMS-based DR-NIR spectroscopic sensor (NeoSpectra-Micro) with advanced machine-learning and deep-learning models, supported by a real-time web interface developed in Gradio for use by surgeons during procedures. By analyzing optical interactions with key tissue chromophores—hemoglobin, water, and lipids—the platform identifies distinctive spectral signatures of malignancy and generates spatially resolved tumor-margin maps. In addition, the study introduces a novel Metric for Classification Assessment and Success (MCAS), a clinically weighted evaluation metric that emphasizes diagnostic safety by balancing sensitivity and specificity according to medical priorities.

Ultimately, this work aims to establish a portable, AI-enhanced NIR spectroscopy system that delivers rapid, accurate, and interpretable feedback within the time frame of standard surgical procedures. The integration of compact optical hardware, data-driven intelligence, and real-time visualization represents a critical step toward practical, intraoperative adoption of label-free cancer margin assessment, improving surgical precision and patient outcomes.