Scholarlink

<p>Kidney disease identification is crucial in medical diagnostics, necessitating rapid and precise models that can detect anomalies in medical imaging. This paper introduces the KiCNN model, a lightweight and efficient convolutional neural network (CNN) combined with a Spatial Attention Mechanism (SAM) to improve feature extraction from four types (Normal, Cyst, Stone, and Tumour) of 12,446 CT scans. Similarly, MobileNet, Xception, and ResNet50 were enhanced with SAM. Performance was evaluated using various combinations of pooling layers (Max Pooling and Average Pooling) and optimizers (Adam and RMSprop). KiCNN demonstrated a 4-fold average Matthews correlation coefficient (MCC) and kappa score (KS) of ~ 100.0%, whereas MobileNet, Xception, and ResNet50 with SAM obtained MCC of 99.9%, 99.5%, and 99.8%, respectively. This outcome demonstrated that the SAM greatly enhanced feature representation by emphasizing spatially important areas in CT scans. Compared to all models, KiCNN was small, with 396.5KB, 101,504 parameters, and 17 layers, including four convolutional layers (CLs). The findings highlight the KiCNN as a lightweight solution that may be deployed on resource-constrained devices and in the cloud. This work demonstrates the utility of customized CNN models and SAM-enhanced TL structures for effective kidney disease identification, paving the way for practical clinical applications. Despite this performance, numerous obstacles exist, including the lack of patient-level metadata. Additionally, the dataset is in a controlled environment, which limits its generalizability to practical healthcare environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

KiCNN: A lightweight CNN model for kidney disease detection using spatial attention mechanisms

  • 期刊论文

  • 发表日期：

    2026年3月15日

查看全文
  • Kamini G. Panchbhai,
  • Madhusudan G. Lanjewar

摘要

Kidney disease identification is crucial in medical diagnostics, necessitating rapid and precise models that can detect anomalies in medical imaging. This paper introduces the KiCNN model, a lightweight and efficient convolutional neural network (CNN) combined with a Spatial Attention Mechanism (SAM) to improve feature extraction from four types (Normal, Cyst, Stone, and Tumour) of 12,446 CT scans. Similarly, MobileNet, Xception, and ResNet50 were enhanced with SAM. Performance was evaluated using various combinations of pooling layers (Max Pooling and Average Pooling) and optimizers (Adam and RMSprop). KiCNN demonstrated a 4-fold average Matthews correlation coefficient (MCC) and kappa score (KS) of ~ 100.0%, whereas MobileNet, Xception, and ResNet50 with SAM obtained MCC of 99.9%, 99.5%, and 99.8%, respectively. This outcome demonstrated that the SAM greatly enhanced feature representation by emphasizing spatially important areas in CT scans. Compared to all models, KiCNN was small, with 396.5KB, 101,504 parameters, and 17 layers, including four convolutional layers (CLs). The findings highlight the KiCNN as a lightweight solution that may be deployed on resource-constrained devices and in the cloud. This work demonstrates the utility of customized CNN models and SAM-enhanced TL structures for effective kidney disease identification, paving the way for practical clinical applications. Despite this performance, numerous obstacles exist, including the lack of patient-level metadata. Additionally, the dataset is in a controlled environment, which limits its generalizability to practical healthcare environments.