Authors - Arathi B K, Rishikeshwar Kumaresan, S Kanagalakshmi, Sathish Kumar S Abstract - Single magnetic resonance imaging (MRI) super‑resolution remains challenging due to the substantial heterogeneity between low‑ and high‑resolution (LR-SR) inputs. This paper presents an ablation analysis of three convolutional neural‑network architectures, namely Conv2D, fully convolutional network (FCN), and U‑Net, combined with four activation functions (Linear, Tanh, ReLU, Leaky ReLU). LR inputs are generated through mean- and max‑pooling with a 6×6 scale factor, enabling evaluation under both smooth and heterogeneous degradation conditions. The results show that U‑Net achieves the highest reconstruction accuracy, reducing MAE by 8% relative to FCN and 10% relative to Conv2D. ReLU-based activations provide stable convergence for shallow models, while the U-Net remains robust across all activation functions. These findings emphasise the importance of selecting appropriate architectures and activation functions to achieve robust and high‑quality MRI super‑resolution in real‑world applications.
Open Zoom