BREAST-RANKNet: a fuzzy rank-based ensemble of CNNs with residual learning for enhanced breast cancer detection from ultrasound and mammogram images

Abstract

Abstract Breast cancer (BC) detection with medical imaging, like ultrasound and mammography, has advanced with deep learning (DL). However, static ensemble models struggle to adapt to varying classifier confidence, and traditional methods fail to capture both local and global features effectively. In this paper, we introduce a novel approach, BREAST-RANKNet, designed to overcome existing limitations by employing an adaptive fuzzy rank-based ensemble strategy. This method dynamically combines the decision scores of three state-of-the-art pre-trained CNN models—DenseNet169, MobileNetV1, and InceptionResNetV2—while accounting for the confidence in the predictions of each model. To enhance the robustness of these base models, we incorporate an Improved Residual Learning Block (IRLB), which integrates depthwise separable convolutions, GELU activations, and residual connections. This block improves computational efficiency and enables the model to capture both local and global features, addressing the challenges posed by complex medical imaging data. Furthermore, we extract probabilities from these models and aggregate them using fuzzy rank-based fusion strategy, utilize three non-linear functions: an exponentially decaying function $$\left( {exp} \right)$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mfenced> <mml:mrow> <mml:mi>exp</mml:mi> </mml:mrow> </mml:mfenced> </mml:math> , the $$tanh$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>tanh</mml:mi> </mml:mrow> </mml:math> function, and the $$sigmoid$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>sigmoid</mml:mi> </mml:mrow> </mml:math> function, allowing for adaptive integration based on model confidence. We enhance the dataset by employing fuzzy contrast enhancement, which improves image quality and ensures better feature extraction for more accurate breast cancer detection. We validate BREAST-RANKNet through extensive experiments, achieving 96.15% on breast ultrasound (BUSI) and 99.45% on a mammography dataset, with additional tests on three datasets, including histopathological and ultrasound images. We use advanced visualization techniques, including Grad-CAM and SHAP, to clarify the model’s focus during classification, enhancing interpretability and reinforcing its applicability in medical diagnostics. The results demonstrate BREAST-RANKNet’s potential as a robust, efficient tool for early BC detection across diverse datasets.

Topics & Concepts