Comparative Analysis of Custom vs. Pre Trained CNN Models for Breast Tumor Classification

Breast tumor classification is crucial for accurate diagnosis and effective treatment planning in breast cancer patients. Traditional diagnostic methods, although informative, are labor-intensive and subject to variability. This study proposes a novel deep learning (DL) framework to enhance breast tumor classification using histopathological images from the BreakHis dataset. The approach involves a comprehensive preprocessing pipeline, including image resizing, data partitioning, and augmentation techniques to improve model robustness. A custom CNN is designed for feature extraction and classification, outperforming pre-trained models such as MobileNetV1, EfficientNetB1, VGG19, and ResNet50V2. The custom CNN achieves a high accuracy of 93.12% and a ROC-AUC of 0.9708. Hyperparameter optimization is performed using Hyperband and Bayesian Optimization, further fine-tuning the model's performance. The Adam optimizer facilitates model convergence during training. Results indicate that the proposed model significantly challenges or exceeds recent state-of-the-art approaches, offering a valuable tool for breast tumor diagnosis and contributing to improved clinical decision-making.