Deep Learning Applications for the Breast Cancer Diagnosis and Classification

Abstract

Breast cancer remains a global health challenge, emphasizing the need for advancements in early detection and diagnosis. High-resolution breast CT scans provide volumetric imaging, offering an advantage over conventional mammography in detecting deep-seated tumors. This study utilizes a dataset of 10,000 CT scans images collected from local hospitals across Pakistan, encompassing diverse demographics. Deep learning-based transfer learning models were applied for tumor classification on unsegmented raw data to access AI’s ability to generalize in real-world scenarios. Among tested models, VGG16 achieved the highest accuracy of 82%, whereas deeper architectures such as ResNet50V2 and EfficientNetB0-B5 struggled with generalization, achieving only 50% accuracy and high false positive rate. To address these limitations, instance segmentation was employed using Roboflow, allowing precise tumor region extraction. YOLOv11 model was then applied to segmented dataset, achieving precision of 88%, recall of 0.90. Loss functions converged effectively, confirming model stability. This approach enhances real-time tumor detection, improving early screening and diagnosis. Future research should optimize segmentation strategies on more and diverse dataset and enhance model performance to deploy model in various clinical setups.