Table 6 Hyperparameters used in SAE and their explanations
From: A lightweight CVTC model for accurate Alzheimer’s MRI analysis and lesion annotation
Parameter | Value | Description |
|---|---|---|
Convolutional Kernel Sizes | Layer 1: (2, 4, 6, 8) Layer 2: (2, 4) Layer 3: (2, 4) Layer 4: (2, 4) | Sorted in descending order for multi-scale feature extraction; larger kernels capture global patterns, while smaller kernels capture local details. |
Output dimension allocation | \([\frac{d}{2},\frac{d}{4},\frac{d}{8},1-\frac{7d}{8}]\) | Sum equals the output dimension d; larger kernels are assigned fewer dimensions to balance computational load. |
Stride | 2 | Stride for each convolutional layer, used to halve the spatial dimensions. |
Padding | \(\frac{(kernel-stride)}{2}\) | Calculated to ensure output dimensions are half of the input dimensions. |
Activation function | ReLU | Applied after each convolution. |
Normalization | Layer normalization (eps = 1e-5) | Stabilizes gradient flow. |
