Pneumonia Detection AI: High-Recall Medical Imaging

In medical screening, the cost of a "False Negative" is infinitely higher than a "False Positive". For our Pneumonia Detection AI, the primary objective was Medical Recall — to ensure that if pneumonia was present in a chest X-ray, the model would catch it 99% of the time.

I designed a custom CNN with four primary convolutional blocks, each utilizing Batch Normalization and Max-Pooling. We used Global Average Pooling to reduce overfitting. To make the model robust, I implemented rigorous data augmentation including horizontal flips, rotations (up to 20 degrees), and brightness adjustments.

To solve the "Black Box" problem, I integrated Grad-CAM. It produces a heatmap overlaid on the original X-ray, allowing physicians to see exactly where the model is looking. The model achieved 99% Medical Recall, missing only 3 cases out of 390 in the test set.