Breast Cancer Ultrasound Image Classification: Comparative Study and Practice of Three Neural Network Paradigms

This article provides an in-depth analysis of an open-source project that compares the performance of traditional artificial neural networks (ANN), spiking neural networks (SNN), and physics-informed neural networks (PINN) in breast cancer ultrasound image classification tasks. It explores the application value of different architectures in medical image diagnosis and provides references for technology selection.

Artificial intelligence is rapidly developing in the field of medical image diagnosis. As a common malignant tumor in women, early screening for breast cancer is a research focus. Ultrasound examination is non-invasive, radiation-free, and low-cost, but its interpretation relies on doctors' experience, and the features of benign and malignant lesions overlap. AI-assisted systems are expected to improve diagnostic accuracy and consistency.

The project compares three paradigms: 1. Traditional ANN (based on backpropagation, mature and stable but requires large amounts of labeled data and has high energy consumption); 2. SNN (bionic spiking mechanism, event-driven with low energy consumption but complex training); 3. PINN (incorporates physical constraints, improves generalization and interpretability but has difficulty in constraint selection). The BreastMNIST dataset is used, and evaluation metrics include accuracy, sensitivity, specificity, and AUC-ROC, with a focus on the trade-off between sensitivity and specificity in clinical scenarios.

Traditional ANN is mature and supports transfer learning but has high data requirements; SNN has low energy consumption, is suitable for edge devices, and has potential for temporal processing; PINN can integrate domain knowledge to improve interpretability but requires balanced constraint design. Recommendations for selection in different scenarios: choose SNN for resource-constrained environments, PINN for interpretability needs, and optimized ANN for accuracy pursuit.

Applications need to focus on data privacy protection, algorithm fairness (avoiding group bias), and clinical integration (integrating into diagnosis and treatment processes, handling disagreements between AI and doctors' judgments). These issues require collaboration among technical developers, clinicians, ethicists, and other parties to resolve.

Future medical AI may integrate the advantages of multiple paradigms to build more robust, efficient, and interpretable systems. This open-source project provides code implementations and a framework for thinking about technology selection, which is of reference value to researchers and developers.