Enhanced Mammographic Tumor Detection Through AI Based Techniques

Abstract

Mammography for breast cancer screening is essential for early detection, but analyzing these images is difficult because of subtle visual differences among radiologists and the labor-intensive process of manual evaluation. To overcome these limitations, an explainable deep learning framework is pro-posed in this paper, which not only automatically performs benign/malignant classification on mammographic images but also maintains high transparency during the decision-making procedure. The proposed work is a 5-layer CNN model of dense layers with two layers at the top with a VGG16 which is a popular pre-trained model.

Distinctive methods, like flipping, small junctions, versions, rocketing, and contrast changes effect in a robust model. An upfront Flask front-end serves real-time interpretation with a heatmap overlay and novel processed images side-by-side, creating the pipeline useful for auxiliary clinical system as well as educational purposes. We gauge the model by observing correctness, ROC-AUC, precision recall AUC, sensitivity, specificity, confusion matrices, calibration curves, as well as prediction confidence to identify borderline instances-es.

The proposed model is considered to be clear — Integrated Gradients stretch pixel-level attribution and Grad-CAM provides abrasive localization to ensure predictions align with clinically relevant patterns. Results demonstrate that the light-weight CNN can be comparable to VGG16 when the preprocessing and augmentation are suitably done. A simple, modular approach has improved radiologists’' decision-making and confidence in AI-enabled breast imaging—consistent with clinical practice and scalable to include multi-view fusion, lesion localization, and triage.