EffiRadNet: Lightweight and User-Friendly Open-Source EfficientNet-Based Model for Radiology Image Binary Classification Tasks

Objectives To develop and evaluate EffiRadNet, a lightweight, user-friendly, open-source image classification model based on EfficientNet-B0, tailored for simple binary classification tasks in radiology, involving visually distinct image. The model aims to provide high accuracy and usability while remaining computationally efficient and accessible to non-engineering users. Methods EffiRadNet was trained and validated on three task-specific, class-balanced datasets (200 training and 200 validation images per dataset) sourced from the Open Access Biomedical Image Search Engine. The three classification tasks included a “Medical” task distinguishing radiological from non-radiological medical images, a “Modality” task classifying X-rays versus other radological modalities, and a “X-rays” task differentiating chest X-rays from other anatomical X-rays. The model used transfer learning with EfficientNet-B0 pretrained on ImageNet, fine-tuned with a modified two-class output layer. Performance was assessed across different hyperparameter settings using AUC, sensitivity, specificity, positive predictive value and negative predictive value. Results EffiRadNet achieved strong classification performance across all tasks, with AUC values up to 1.0 under optimal hyperparameters (≥≤30 epochs, batch size = 16, learning rate = 0.0001). The “Medical” and “X-rays” models showed balanced sensitivity and specificity, while the “Modality” model displayed high sensitivity but variable specificity depending on hyperparameters. Training times ranged from 11 to 12 minutes, inference took less than 30 seconds per dataset. Conclusion EffiRadNet is a fast, accurate, and accessible AI tool for binary classification of radiological images. Its open-source availability and minimal hardware requirements make it well-suited for tasks such as image modality classification, dataset preprocessing, and quality control.