CXR AI Model Configuration: ResNet-18 for 3-Class Pneumonia Detection

Alejandro Hernández-Arango published a dataset containing the configuration for a chest X-ray AI model on 2026-04-30. The model uses a ResNet-18 architecture, pre-trained on multiple public datasets and fine-tuned on 395 images from Clínica CES. The configuration includes preprocessing steps, training hyperparameters, and details for generating Class Activation Maps.

Use Cases

• Reproducing a ResNet-18-based pneumonia detection model based on the specified architecture and hyperparameters.
• Analyzing the impact of preprocessing steps like CLAHE and lung segmentation on model performance.
• Comparing training strategies using the detailed hyperparameters, including AdamW optimizer and cosine annealing with warm restarts.
• Investigating model interpretability using the described Class Activation Map (CAM) generation method.

Strengths

• Model architecture and training hyperparameters are explicitly detailed, including 300 epochs, AdamW with a learning rate of 0.005, and a batch size of 512.
• Preprocessing pipeline is clearly described, involving CLAHE and DeepLabV3 lung segmentation.
• Pre-training leveraged six established public datasets (CheXpert, NIH ChestX-ray14, etc.), and fine-tuning used 395 images from a specific clinical source.

Limitations

• The dataset is very small at 6.2 KB, suggesting it contains only configuration metadata, not the actual image data or model weights.
• Row count and column-level documentation are absent; the structure and semantics of the XLSX file must be inferred after download.
• The fine-tuning dataset of 395 images from a single clinic may limit the generalizability of the model configuration.

Provenance

Source

Alejandro Hernández-Arango via figshare.

Collection Method

Likely compiled from the development and training process of a specific AI model.

Freshness

Last updated 2026-04-30 17:49:49; freshness should be verified.

Geography

Fine-tuning data sourced from Clínica CES; specific location is unknown.