Convolutional Fourier Analysis Network (CONV-FAN-POX): A novel Time-Frequency approach for medical image analysis

The recent monkeypox outbreak underscores the urgent need for improved diagnostic tools to enable swift, accurate disease detection. This study introduces CONV-FAN-POX, a novel neural architecture that integrates Fourier Analysis Network (FAN) principles into a deep learning pipeline for medical image classification. Unlike traditional models that rely solely on spatial-domain features, FAN leverages the Fourier series to enable robust modeling of quasi-periodic patterns and to capture global frequency-domain characteristics essential for accurate diagnosis. The proposed model, utilizing an EfficientNetV2 backbone, was evaluated on the MSLD2.0 dataset, achieving an average accuracy of 0.9881 and an F1-score of 0.9856 across five-fold cross-validation in a six-class setting. To validate the model's robustness and the specific contributions of the frequency-domain approach, extensive ablation studies were conducted, including a direct comparison of FAN and Dense-layer architectures, training and evaluation on the MSID dataset, and cross-dataset transfer testing. Furthermore, Explainable AI (XAI) was applied using Grad-CAM, which provides visual evidence that the FAN layer effectively prioritizes lesion-relevant features over background noise. Achieving superior performance while utilizing significantly fewer trainable parameters than conventional architectures, these results highlight CONV-FAN-POX as an efficient, interpretable, and generalizable alternative for medical image analysis, particularly in time-frequency-rich diagnostic contexts.