ANALYSIS OF NEURAL MODEL EFFECTIVENESS IN ULTRASOUND DIAGNOSTIC IMAGE PROCESSING

Abstract

This article presents a comparative analysis of the effectiveness of using simple artificial neural networks (ANN) and convolutional neural networks (CNN) for processing ultrasound diagnostic images. The relevance of the study stems from the widespread use of ultrasound imaging in clinical practice, along with its inherent limitations such as low contrast, high noise levels, and dependency on both the equipment and operator’s experience. In the context of increasing volumes of medical data and the growing need for automated decision support systems, neural models offer a promising tool for enhancing the diagnostic value of ultrasound images. To achieve the research objective, an experimental simulation was conducted using an open-access dataset of breast ultrasound images available on the Kaggle platform. The original images were downscaled from 500 × 500 to 128 × 128 pixels to reduce computational load. Two neural network architectures were implemented: a basic ANN with a single hidden layer and a CNN with three convolutional layers. The models were evaluated based on classification accuracy, recall, precision, F1-score, inference speed, and interpretability.The results demonstrated that CNN significantly outperformed ANN across all major performance metrics. CNN achieved an average classification accuracy of 86.4 %, compared to 71.3 % for ANN. Furthermore, CNN provided more stable performance under varying image quality and noise conditions and enabled the use of Grad-CAM for highlighting the most influential regions in the decision-making process. While ANN was less sensitive to local textural features, it offered advantages in computational simplicity and inference speed (under 5 ms per image).The study concludes that CNN is more suitable for clinical settings where accuracy and reliability are critical, whereas ANN may be useful for preliminary filtering or in systems with limited computational resources. The article also emphasizes the importance of considering hardware-specific variations in image sources and highlights the future potential of hybrid, interpretable neural architectures in medical imaging.