This investigation presents a specialized multimodal neural network for classifying image-based electrocardiogram exams (ECG) to distinguish between normal cardiac rhythms and Atrial Fibrillation (AF) using a dataset exclusively comprising ECG exam images. The model exhibits a preprocessing stage adept at extracting the DII lead from PNG images. Subsequently, we use the extracted lead to generate a time series and a spectrogram input to feed the multimodal network. The cross-validation metrics demonstrate the efficacy of the methodology with an accuracy of 97.65%, AUC of 94.08%, specificity of 96.89%, sensitivity of 99.20%, and an F1 score of 96.57%. Additionally, the methodology exhibits impressive performance across various data sources and multiple folds, achieving an average accuracy of 90.70%, AUC of 90.78%, specificity of 90.62%, sensitivity of 90.94%, and an F1 score of 82.09%. The multimodal approach recommended here eliminates the need for specialized software, making it easier to integrate into clinical practice and enhancing the diagnostic capabilities of healthcare professionals. Our multimodal approach enhances the clinical relevance of AF detection, utilizing image ECG exams commonly used in real-world scenarios. This innovative methodology incorporates information from images, spectrograms, and time series derived from the original ECG exams, providing a comprehensive analysis.