Abstract:
It is a grave health problem on a global scale and diagnosis of breast cancer should be evident at the initial level, which will enhance results. This article creates effective deep learning systems on the BreakHis dataset of binary classification of breast histopathological cultures. To select the best model in benign, MobileNet, VGG, Inception and ResNet will be used and trained and tested several CNN models including EfficientNet to classify the training benign tissue, malignant tissue and classify benchmark tissue. Both the base models and fine-tuned models are investigated to establish the impacts of unfreezing of the weight and changing features on performance. We also prepare hybrid/ensemble models including EfficientNetB0.+ MobileNetV2 to use complementary features of the two networks to improve the accuracy of classification. The workflow involves preprocessing the data, training and building of the model and the model performance in terms of accuracy, precision, recall, F1 score and loss. We examine energy usage and processing overhead to determine whether they can be implemented in low-resource settings or in the clinical context. Findings indicate that fine-tuning could be used to attain a significant performance enhancement and the proposed hybrid model leads to a greater accuracy at the affordable cost. It is also estimated by the study that increasing the magnification of images resulted in the model generalizing more. Taking into consideration these weaknesses, we observe that the collective deep learning technique on breast cancer is required that is effective to design networks and to be integrated in future both in the clinic and in multi-modal systems.
