Pneumonia from chest radiographs, and what recall reveals
Transfer learning against training from scratch on chest X-rays, where per class recall shows which way the model leans.
82%
Test accuracy
pretrained
0.99
Pneumonia recall
0.55
Normal recall
where the model leans
0.96
Normal precision
The question
The dataset is heavily skewed toward pneumonia cases, so a model can score well by leaning on the majority class. In a screening setting the asymmetry matters enormously: a missed pneumonia carries very different weight from a false alarm.
Approach
ResNet-34 trained two ways, once initialised from ImageNet weights and once from scratch, with everything else held constant. I reported per class precision and recall alongside accuracy, plus confusion matrices and a gallery of the misclassified images.
Outcome
The pretrained model reached 82 percent test accuracy, with precision 0.96 and recall 0.55 on normal cases and precision 0.78 with recall 0.99 on pneumonia. Trained from scratch, recall on normal fell to 0.21 while pneumonia recall reached 1.00. Both models leaned heavily toward calling a case pneumonia, and the scratch model leaned further.
What it taught me
Looking at the misclassified images was worth more than any metric. Reading the individual mistakes is how you find out whether the model learned the pathology or the acquisition protocol.