MSA-ENet Multi-Scale Attention Enhanced Network for Kidney Stone Detection in Ultrasound Images

Project Code :TCMAPY2314

Objective

The objective of this project is to develop an advanced deep learning model, MSA-ENet, for the efficient and accurate detection of kidney stones in ultrasound images. The model aims to leverage multi-scale feature extraction and dual attention mechanisms to enhance detection performance, especially for stones of varying sizes. By utilizing EfficientNet-B0 as the backbone, MSA-ENet extracts hierarchical feature maps across different spatial resolutions. The model also integrates ensemble learning with three classifiers to improve robustness and generalization. A custom composite loss function is used to address class imbalance and reduce overfitting. The primary goal is to achieve high accuracy and reliability in kidney stone detection, outperforming traditional CNN approaches. Additionally, the integration of Grad-CAM aims to enhance model interpretability, making it suitable for real-world clinical deployment.

Abstract

This work presents MSA-ENet, a deep learning model designed for the detection of kidney stones in ultrasound images. The architecture incorporates multi-scale feature extraction, dual attention mechanisms, and ensemble learning to achieve high performance in classification tasks. At its core, EfficientNet-B0 serves as the backbone, extracting hierarchical feature maps across various spatial resolutions, allowing the detection of kidney stones in different sizes. The Multi-Scale Feature Extraction (MSFE) module processes these feature maps using multiple parallel branches to capture both fine-grained details and broader contextual patterns. The dual attention mechanism, including channel and spatial attention, enhances the focus on significant features and spatial areas for improved detection. Additionally, MSA-ENet uses an ensemble of three classifiers, which further strengthens its prediction robustness and generalization capability. A custom loss function, which combines focal loss, label smoothing, and cross-entropy, helps address class imbalance and prevents overfitting. The model has been evaluated on two datasets: the Kidney Stone Ultrasound Dataset from Roboflow and the Kidney Stone Classification and Object Detection Dataset from Kaggle, achieving accuracies of 98.25% and 98.94%, respectively, with ROC-AUC scores of 0.9935 and 0.9999. When compared to traditional CNN models, MSA-ENet demonstrates superior performance in accuracy, robustness, and generalization. Furthermore, the use of Grad-CAM enhances interpretability, making MSA-ENet a strong candidate for clinical application. This study highlights the effectiveness of multi-scale and attention-based architectures in medical image analysis, particularly for kidney stone detection.

Keywords: Kidney stone detection, Ultrasound imaging, MSA-ENet, Multi-scale feature extraction, Dual attention mechanism, EfficientNet-B0, Ensemble learning, Grad-CAM, Medical image analysis, Deep learning, Object detection, Class imbalance, Clinical deployment.

NOTE: Without the concern of our team, please don't submit to the college. This Abstract varies based on student requirements.