Drug-induced liver injury (DILI) remains a major challenge in pharmaceutical research and clinical practice, frequently resulting in drug withdrawal and costly delays in development. Early identification of potential hepatotoxicity is crucial for patient safety and the success of drug candidates. In this paper, we introduce ResDILI, a deep learning–based framework that applies ResNet-18 and ResNet-34 architectures to predict DILI directly from molecular structure images. ResDILI leverages the power of deep learning and the ResNet architecture to extract meaningful features from chemical structure images and predict the likelihood of DILI occurrence. Our approach utilizes a large dataset of 7054 diverse drug compounds that were carefully curated from previous studies and annotated with comprehensive liver injury information. Using stratified 10-fold cross-validation, ResDILI achieved notable predictive performance with an accuracy of 0.901, area under the ROC curve of 0.951, sensitivity of 0.911, specificity of 0.885, and Matthews correlation coefficient of 0.821. Better performance than the existing approach when comparing evaluations on external datasets. This model contributes to the advancement of DILI prediction methods and holds promise for early-stage assessment of drug safety in modern drug discovery. However, the model still relies on public data and needs further validation on experimental and clinical datasets to confirm its robustness.