Table 1 Architectural breakdown of the proposed model.
From: Trustworthy pneumonia detection in chest X-ray imaging through attention-guided deep learning
Block | Role | Key parameters | Constitution (components) |
|---|---|---|---|
Image input | Accepts image input | 3 × 128 × 128 (RGB) | Input tensor of shape (batch_size, 3, 128, 128) |
Spatial feature extraction block | Extracts spatial features | Conv2d (3 → 32), Conv2d (32 → 64), Conv2d (64 → 128) | Kernel = 3 × 3, Stride = 1, Padding = 1, BatchNorm, LeakyReLU, MaxPool (2 × 2, Stride = 2) |
Flatten | Converts CNN features into vector | Flatten to 1D | Converts CNN output to a vectorized format |
Temporal dynamics modeling block | Captures temporal dependencies | num_steps = 25, BiGRU = 2, Hidden Size = 256, Dropout = 0.3 | Bidirectional GRU layers |
Feature projection linear | Ensures smooth CNN → SNN transition | fc1 (65,536 → 128), fc1 (128 → 64) | Fully connected layers |
Spiking neural processing block | Processes spikes across time | num_steps = 25, Spike Aggregation = Mean | Time steps, spike aggregation |
Decision head block | Outputs final classification | Fc1 (768 → 512), Dropout = 0.4 Fc2 (512 → 128), Dropout = 0.4 Fc3 (128 → 2) | Fully connected Layers + Dropout |
