The PhysioNet/Computing in Cardiology Challenge 2021 focused on exploring the utility of reduced-lead ECGs for arrhythmia classification. We proposed a novel multi-scale convolutional neural network that can classify 30 scored arrhythmias from 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead ECGs. The proposed network was achieved by multiple branch networks to effectively extract the pathological information from micro-scale to macro-scale via convolution kernels (filters). The backbone of each branch network was a carefully designed 2-D convolutional neural network (CNN) with residual connection and attention mechanism, and it can adapt to multi-lead ECGs as input. The first 10 seconds of records from corresponding leads were extracted and preprocessed as input for end-to-end training, and the prediction probabilities of 30 categories were output. The proposed algorithms were firstly evaluated on officially published datasets of 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead ECGs, and we achieved 5-fold cross-validation scores of 0.719, 0.687, 0.698, 0.695, and 0.678 by using the challenge metric. Finally, our team, AIRCAS_MEL1, achieved challenge validation scores of 0.63, 0.57, 0.58, 0.57, and 0.56 for 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead ECGs, respectively. Results showed that reduced-lead ECGs can indeed capture a wide range of arrhythmia diagnostic information, and some (4-lead, 3-lead) are even comparable to 12-lead ECGs in some limited contexts. Those reduced-lead ECGs (4-lead, 3-lead) are promising in developing smaller, lower-cost, and easier-to-use diagnostic devices that are comparable to the 12-lead standard diagnostic system.