Interpretable Deep Learning for Pediatric Pneumonia Diagnosis Through Multi-Phase Feature Learning and Activation Patterns DOI Open Access

Petra Radočaj,

Goran Martinović

Electronics, Год журнала: 2025, Номер 14(9), С. 1899 - 1899

Опубликована: Май 7, 2025

Pediatric pneumonia remains a critical global health challenge requiring accurate and interpretable diagnostic solutions. Although deep learning has shown potential for recognition on chest X-ray images, gaps persist in understanding model interpretability feature during training. We evaluated four convolutional neural network (CNN) architectures, i.e., InceptionV3, InceptionResNetV2, DenseNet201, MobileNetV2, using three approaches—standard convolution, multi-scale strided convolution—all incorporating the Mish activation function. Among tested models, with convolutions, demonstrated best performance, achieving an accuracy of 0.9718. InceptionV3 also performed well same approach, 0.9684. For DenseNet201 convolution approach was more effective, accuracies 0.9676 0.9437, respectively. Gradient-weighted class mapping (Grad-CAM) visualizations provided insights, e.g., convolutions identified diffuse viral patterns across wider lung regions, while precisely highlighted localized bacterial consolidations, aligning radiologists’ priorities. These findings establish following architectural guidelines: are suited to hierarchical CNNs, approaches optimize compact models. This research significantly advances development interpretable, high-performance systems pediatric X-rays, bridging gap between computational innovation clinical application.

Язык: Английский

Interpretable Deep Learning for Pediatric Pneumonia Diagnosis Through Multi-Phase Feature Learning and Activation Patterns DOI Open Access

Petra Radočaj,

Goran Martinović

Electronics, Год журнала: 2025, Номер 14(9), С. 1899 - 1899

Опубликована: Май 7, 2025

Pediatric pneumonia remains a critical global health challenge requiring accurate and interpretable diagnostic solutions. Although deep learning has shown potential for recognition on chest X-ray images, gaps persist in understanding model interpretability feature during training. We evaluated four convolutional neural network (CNN) architectures, i.e., InceptionV3, InceptionResNetV2, DenseNet201, MobileNetV2, using three approaches—standard convolution, multi-scale strided convolution—all incorporating the Mish activation function. Among tested models, with convolutions, demonstrated best performance, achieving an accuracy of 0.9718. InceptionV3 also performed well same approach, 0.9684. For DenseNet201 convolution approach was more effective, accuracies 0.9676 0.9437, respectively. Gradient-weighted class mapping (Grad-CAM) visualizations provided insights, e.g., convolutions identified diffuse viral patterns across wider lung regions, while precisely highlighted localized bacterial consolidations, aligning radiologists’ priorities. These findings establish following architectural guidelines: are suited to hierarchical CNNs, approaches optimize compact models. This research significantly advances development interpretable, high-performance systems pediatric X-rays, bridging gap between computational innovation clinical application.

Язык: Английский

Процитировано

0