Swarm-driven fine-tuned hybrid transfer learningarchitecture on breast ultrasound scans for advanced prognostic

Breast Malignancy (Tumor) presents a serious health risk as it is the primary cause of cancer-related death for women globally. To enhance the outcomes for patients and ensure optimal treatment, early testing and diagnosing are important. In recent years, deep learning (DL) techniques have shown great promise in aiding the detection and diagnosis of breast maliganacy. In this study, the Breast Echography scans were taken from Breast Ultrasound Scans (BUS) dataset and were resized into 224 × 224, and finally, data pre-processing steps were used. Next, the collective and iterative nature of Particle swarm optimization (PSO), where the swarm of particles collaboratively refined the solution space, to enables the dynamic convergence towards an optimal solution. Further, five transfer learning models were implemented for a 3-class Breast Malignancy classification task ResNet50, EfficientNetB4, MobileNetV2, and InceptionV3. We propose a hybrid model that integrates the three top-performing Transfer Learning (TL) architectures: EfficientNetB4 + MobileNetV2 + InceptionV3. These models’ outputs are combined into a single representation by flattening and concatenating them. To maximize the weight learning of the integrated models, deep dense layers (DDL) are then applied to this combined output. The proposed hybrid approach effectively categorizes medical images related to Breast Malignancyas either malignant, benign, or pertaining to healthy individuals without prior knowledge of the presence of a cancerous lesion. From the experimental evaluation, it is shown that the proposed novel stacking model gives the highest Accuracy of 95.92%, Precision of 95%, Recall of 93%, and F1 Score of 98.12%.