Skip to content

MODEL_CARD.md

Latest commit

 

History

History
116 lines (83 loc) · 4.47 KB

MODEL_CARD.md

File metadata and controls

116 lines (83 loc) · 4.47 KB
Raw

Model Card — DenseNet121 Pneumonia Classifier

Model Details

Field Value
Model Name DenseNet121 Chest X-Ray Classifier
Architecture DenseNet-121 (Huang et al., 2017)
Framework PyTorch / TorchVision
Pre-training ImageNet-1K
Fine-tuning Kaggle Pediatric Pneumonia Dataset
Output Classes 3 (Bacteria, Normal, Virus)
Input Size 224 × 224 × 3 (RGB)
Parameters ~7M
File Size ~27 MB (.pth)

Intended Use

  • Primary Use: Educational tool for understanding AI-assisted medical imaging
  • Target Users: Students, researchers, and ML practitioners
  • Out-of-Scope: Clinical diagnosis, patient treatment decisions, regulatory submissions

Training Data

Property Detail
Dataset Chest X-Ray Images (Pneumonia)
Source Guangzhou Women and Children's Medical Center
Population Pediatric patients (ages 1–5)
Split Method Patient-level (80% train / 10% val / 10% test)
Augmentation Random horizontal flip, rotation (±10°), brightness/contrast jitter

Class Distribution (approximate)

Class Count Percentage
Bacteria ~2,500 ~50%
Normal ~1,300 ~26%
Virus ~1,200 ~24%

Note: Class imbalance is addressed using Weighted Cross-Entropy Loss.

Performance Metrics

Evaluated on the held-out test set (patient-level split):

Metric Value
Infection Sensitivity 98.6% (only 6/433 sick patients missed)
Normal Accuracy 95.3% (162/170)

Confusion Matrix Summary

Pred: Bacteria Pred: Normal Pred: Virus
Actual Bacteria 245 3 52
Actual Normal 0 162 8
Actual Virus 20 3 110

Key Observations

  • Strength: Near-zero false negatives for pneumonia detection (high recall)
  • Weakness: Bacteria vs. Virus confusion (52 Bacteria predicted as Virus), which is expected since radiological features overlap significantly
  • Clinical Safety: The model is biased toward detecting disease, minimizing the risk of sending sick patients home

Explainability

Grad-CAM heatmaps confirm the model focuses on:

  • Lung parenchyma (the tissue where pneumonia manifests)
  • Areas of consolidation and opacification

The model does not attend to:

  • Image borders, text labels, or bone structures
  • This indicates the model has learned medically relevant features

Limitations

  1. Pediatric Only: Trained exclusively on pediatric X-rays (ages 1–5). While the underlying radiological features of pneumonia (consolidation, opacities) are similar across age groups, the model has not been validated on adult data. A separate adult model using the NIH ChestX-ray14 dataset is available (see Future Work below).
  2. Single Dataset: No external validation on other hospital datasets
  3. Binary Subtypes: Only distinguishes Bacterial vs. Viral — does not identify specific organisms
  4. Image Quality: Performance may degrade on low-resolution, rotated, or cropped images
  5. Not FDA Approved: This is a research tool, not a medical device

Future Work — Adult Model (NIH ChestX-ray14)

To address Limitation #1, a second model has been developed using the NIH ChestX-ray14 adult dataset:

Property Pediatric Model Adult Model
Dataset Kaggle Pediatric Pneumonia NIH ChestX-ray14
Population Ages 1–5 Adult patients
Classes 3 (Bacteria, Normal, Virus) 2 (Pneumonia, Normal)
Model File densenet121_pneumonia.pth densenet121_nih_pneumonia.pth
Notebook Colab_Model_training.ipynb NIH_Adult_Training.ipynb

Note: The NIH dataset does not distinguish between Bacterial and Viral pneumonia, so the adult model is a binary classifier (Pneumonia vs Normal).

Ethical Considerations

  • The model should never be the sole basis for clinical decisions
  • False negatives (missed pneumonia) could have serious consequences in real settings
  • The dataset represents a single geographic population; bias against other demographics is possible
  • Any deployment in clinical settings would require extensive validation and regulatory approval