CVV_15M_SARS-CoV-2 is a convolutional neural network to classify chest X-ray images with high accuracy on mobile hardware. Finely tuned for performance on Apple M-series CPUs.
The implementation incorporates ReLU activation functions, LRN, overlapping pooling, and dropout layers. The model can extract intricate patterns from X-ray images at a high degree of accuracy.
TensorFlow employs ReLU activations:
\[ f(x) = \max(0, x) \]
ReLU accelerates trainingon smaller datasets by preventing vanishing gradients and promoting sparse activations.
To enhance feature selectivity, our model implements LRN which ecourages higher competition among neurons:
\[ b_{i,x,y} = \frac{a_{i,x,y}}{\left(k + \alpha \sum_{j=\max(0,i-\frac{n}{2})}^{\min(N-1,i+\frac{n}{2})} (a_{j,x,y})^2 \right)^{\beta}} \]
LRN helps minimize redundancy in datasets with overlapping image features by reducing the impact of already learned features that should be omitted.
By using overlapping pooling regions (\( s < z \)), the model gains a spatial representation of the novelties in the X-rays.
To counter overfitting, dropout (0.5 rate) is applied in the fully connected layers:
\[ P(h_j \mid x) = \sum_i P(h_j \mid i) P(i \mid x) \]
We implemented TensorFlow's mixed precision training policy to optimize computational efficiency:
This approach allows us to:
The model was trained on a dataset of 17,000 chest X-ray images across three categories.
Our implementation leverages a carefully designed structure:
Our model achieved >95% accuracy classifying among three categories:
Image count: 17,000
IMAGE_SIZE: 224
BATCH_SIZE: 16
EPOCHS: 30
NUM_CLASSES: 3
Test accuracy: 0.9403
| Covid | Normal | Pneumonia | |
|---|---|---|---|
| Covid | 708 | 10 | 6 |
| Normal | 70 | 922 | 8 |
| Pneumonia | 11 | 14 | 244 |