Vari Blend Net: An Integrated Multi-Phase Deep Learning Model for Optimizing Breast Cancer Detection from Pre-processing to Explainability

Abstract

Introduction: This study introduces an innovative and comprehensive methodology for detecting breast cancer in histopathology images, utilizing a multi-step approach through Deep Learning (DL) methods to enhance both accuracy and interpretability. The implementation consists of four distinct phases: pre-processing, segmentation, feature extraction, and feature selection which is followed by the detection model, VariBlendNet, which integrates multiple neural network architectures like MobileNet, SqueezeNet, LeNet-5, and GRU with Variation Dropout to capture diverse features and temporal dependencies.

Objectives: To develop a deep learning-based model that can accurately detect and classify breast cancer from histopathological images, addressing the aforementioned challenges to improve diagnostic accuracy and efficiency. To experiment various preprocessing, segmentation, feature extraction and feature selection methods and algorithms gradually to see the effect on accuracy and performance of the model. To generate the fusion of CNN along with the transfer learning concept. To compare the results of proposed work with the results of the research papers based on the same Kaggle dataset.

Methods: The development of the model consists of four distinct phases: pre-processing, segmentation, feature extraction, and feature selection. Amongst which, the data preparation phase consists of Reinhard Colour Normalization ensures consistent colour representation, super pixel-based patch extraction focuses on specific regions for detailed analysis, elastic transformations augment the dataset, and median filtering reduces noise. In the segmentation phase, a U-Net architecture is employed, enhanced with Multi-Head Attention Gates, Focal Loss, and improved skip-connections to accurately identify cancerous regions. The feature extraction process integrates a variety of techniques: deep learning-based features are derived from Inception v3, while texture features are obtained using Local Ternary Patterns (LTP), Gray-Level Difference Statistics (GLDS), and Local Gradient Patterns (LGP). For feature selection, a Hybrid Golden Mongoose Swarm (HGMS) Algorithm, which combines Dwarf Mongoose Optimization and Golden Search, is employed to enhance predictive power. The detection model, VariBlendNet, integrates multiple neural network architectures like MobileNet, SqueezeNet, LeNet-5, and GRU with Variation Dropout to capture diverse features and temporal dependencies.

Results: The model was meticulously designed and trained along with many Preliminary enhancement, image analysis, feature derivation, and feature selection methods as well systematically compared grounded on their performance metrics, including Accuracy, Precision, Recall, and F1 score. The model achieved Accuracy, Precision, Recall and F1 Score values as 99.12%, 98.95%, 98.94% and 98.88% respectively and outperforms all the existing ones.

Conclusions: This holistic methodology aims to advance breast cancer detection by leveraging high-tech techniques in data preparation, image partitioning, attribute extraction, and deep learning-based detection, ultimately contributing to improved diagnostic accuracy 99.12% and thereby interpretability of the model.