Early diagnosis of gastric cancer remains clinically challenging owing to the reliance on invasive procedures such as endoscopy, which often causes patient discomfort and limits widespread screening. Although serum amyloid A has been proposed as a noninvasive biomarker, conventional analysis of mass spectrometry (MS) data involves complex multistep signal processing that introduces cumulative analytical noise. To address this, we developed a method that transforms raw MS data into grayscale images and applies convolutional neural networks to classify samples as gastric cancer, gastritis, or normal. This end-to-end image-based method eliminates the need for baseline correction or peak detection, reducing analytical artifacts and improving reproducibility. The experiments were conducted across two publicly sourced and institutionally collected datasets. The results demonstrated that the method achieved an average accuracy of 0.93 ± 0.05, sensitivity of 0.92 ± 0.06, and specificity of 0.94 ± 0.04, as determined by 10-fold cross-validation. Ablation study results confirmed the importance of each architectural component, while visualization of activation maps verified the biological relevance of spectral regions in this method. These findings underscore the potential of image-based deep learning as a reliable, noninvasive diagnostic tool for gastric cancer screening and proteomic data interpretation.