OPTIMIZED HYBRID MODEL FOR BREAST CANCER CLASSIFICATION: A COMPARATIVE EVALUATION ON PRECISION AND RECALL WITH CONVENTIONAL MODELS

Abstract

Breast cancer is a pressing issue in the global healthcare environment, and proper early detection during the initial stages with the help of mammography is associated with better chances of survival. Nevertheless, low contrast, complicated tissue structures, and inter-observer variability are known to hamper mammogram interpretation. To overcome such obstacles in which this study presents an optimized hybrid deep architecture combining a Transformer encoder, which extracts spatial features, along with a recurrent neural network (RNN) each in the sequence of refining features. The proposed hybrid method, in contrast to the traditional CNN-based systems, which only use local spatial convolution, considers the global contextual dependencies, as well as the intra-image structural dependency, which results in the enhancement of diagnostic accuracy and strength.

The model is tested on CBIS-DDSM mammography data containing around 10,000 high-quality images, after the preprocessing stage, that is normalization, contrast limited adaptive histogram equalization (CLAHE), augmentation, and patch extraction. Accuracy, Precision, Recall, F1-score, MAE, RMSE, and inference latency are some of the performance metrics. The hybrid model has an accuracy of 94.5% and a precision score of 0.947 and a recall score of 0.943, as well as an F1- score of 0.945, MAE of 0.087, and a RMSE of 0.112 which is better than five traditional baselines such as ResNet50, DenseNet121, LSTM-Networks, Vision Transformer (ViT), and MobileNetV3.

The findings show that the sequential learning with spatial attention provides higher sensitivity to minute malignant characteristics. The proposed framework has a high level of performance and low-latency inference (approximately 28 ms/sample) and can be clinically applicable in the context of mammography screening processes. Its application in clouds in the diagnostics systems makes it a potentially useful tool in real-time decision support in screening and triaging of breast cancer.