Pneumonia X-Ray Classifier

EfficientNetB4 · Bacterial vs. Viral · Research Demo

Concat Pooling TF 2.19 Research Only

Fixed Size Ready

384×384 · 32 epochs · AdamW

86.92%Accuracy ℹ

0.9198ROC AUC ℹ

67.57%Viral Recall ℹ

85.02%Macro F1 ℹ

Progressive Resizing Ready

224→256→384 · 41 epochs · AdamW

78.46%Accuracy ℹ

0.8011ROC AUC ℹ

52.70%Viral Recall ℹ

74.72%Macro F1 ℹ

Input

Drop a chest X-ray here

or JPEG · PNG · Pediatric chest X-ray

Upload a chest X-ray to run inference.

Backbone	EfficientNetB4 (ImageNet)
Pooling	Concat (GAP + GMP → 3,584-dim)
Head	Dense 512 · Swish · BN · Dropout 0.5
Output	Dense 2 · Softmax (float32)
Loss	Focal Loss α=0.25, γ=2.0, smooth=0.1
Optimizer	AdamW + LossScaleOptimizer (fp16)
TTA	2-step (original + H-flip)

Train pool	3,883 images (2,538 bact · 1,345 viral)
Test set	N=390 (242 bact · 148 viral)
Split	80 / 20 from train+val folders
Balancing	Dynamic class weights (no oversampling)
Platform	Kaggle · Tesla P100-PCIE-16GB
Framework	TensorFlow 2.19.0 · mixed fp16