MACHINE LEARNING-DRIVEN FRAMEWORK FOR EARLY CANCER DETECTION USING CELLULAR MORPHOMETRIC FEATURES

Abstract

Background: Cancer remains a leading global health challenge, with breast cancer specifically representing a significant portion of oncology caseloads worldwide. Despite advancements in treatment modalities, early detection remains notoriously difficult, and diagnoses often occur at advanced stages where therapeutic efficacy is diminished. Conventional pathological analysis is heavily time-intensive and uniquely subject to inter-observer variability, especially in border-line cases. Artificial Intelligence (AI) and machine learning offer promising pathways to augment clinical accuracy, improve workflow efficiency, and facilitate truly personalized care Objective: To develop, validate, and comprehensively evaluate an interpretable machine learning framework for the binary classification and clinical risk stratification of breast cancer using structured cellular morphometrics. This framework is explicitly designed to serve as a foundational, highly transparent module for a broader multimodal Clinical Decision Support System (CDSS) [2]. Methods: Utilizing the highly validated Breast Cancer Wisconsin (Diagnostic) dataset ( ), four distinct machine learning models (Logistic Regression, Random Forest, Gradient Boosting, and a Multi-Layer Perceptron Neural Network) were trained to classify cytological tumors as benign or malignant. The optimal model was subsequently leveraged to generate continuous probability distributions, establishing actionable, data-driven clinical risk thresholds (Low, Intermediate, High) [5], [6]. Results: All evaluated models demonstrated exceptionally high discriminative performance. Logistic Regression achieved the highest Area Under the Receiver Operating Characteristic Curve (AUROC) at 0.9960, with a sensitivity of 0.9286 and specificity of 0.9861. Feature importance analysis illuminated the biological mechanisms driving the algorithms, revealing that cellular "worst area" and "worst concave points" were the strongest predictors of malignancy. Furthermore, the risk stratification framework successfully separated benign and malignant probability densities, effectively minimizing clinical ambiguity and creating a clear "grey zone" for targeted physician review [7]. Conclusion: The proposed AI-CDSS provides a highly accurate, computationally efficient, and rigorously interpretable method for early breast cancer detection. By providing probabilistic risk stratification rather than mere binary outputs, the system effectively supports clinical triaging and personalized patient management without functioning as an opaque "black box"