Aunsiels's picture
Upload folder using huggingface_hub
f48e315 verified
metadata
library_name: transformers
license: apache-2.0
base_model: microsoft/resnet-50
tags:
  - image-classification
  - pytorch
  - resnet
  - medical
datasets:
  - paultimothymooney/chest-xray-pneumonia
metrics:
  - accuracy
  - f1
model-index:
  - name: resnet-pneumonia-detection
    results:
      - task:
          name: Image Classification
          type: image-classification
        dataset:
          name: imagefolder
          type: imagefolder
          config: default
          split: validation
          args: default
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.833
          - name: F1
            type: f1
            value: 0.835

ResNet-50 for Pneumonia Detection

This model is a fine-tuned version of microsoft/resnet-50 trained to classify chest X-ray images into two categories: NORMAL and PNEUMONIA.

It was trained to handle class imbalances using weighted Cross-Entropy loss and utilizes specific image augmentations suited for medical radiography.

Dataset

The model was trained on the Chest X-Ray Images (Pneumonia) dataset.

  • Normal: 1,341 training images
  • Pneumonia: 3,876 training images

Training Procedure

Preprocessing and Augmentation

Images were resized and normalized using the AutoImageProcessor configuration from the base ResNet-50 model. During training, the following augmentations were applied to improve generalization:

  • Random Resized Crop
  • Random Horizontal Flip
  • Random Rotation (15 degrees)
  • Color Jitter (Brightness and Contrast)

Hyperparameters

  • Learning Rate: 5e-5
  • Train Batch Size: 8
  • Gradient Accumulation Steps: 8 (Effective Batch Size: 64)
  • Eval Batch Size: 8
  • Epochs: 5
  • Warmup Ratio: 0.1
  • Loss Function: Weighted Cross-Entropy Loss (to penalize minority class misclassifications)
  • Optimization Strategy: Best model loaded at the end based on F1 score.

Evaluation Results

On the standard test split, the model achieved the following performance:

  • Accuracy: 0.833
  • F1 Score (Weighted): 0.835