A classifier can look accurate while doing a poor job of finding what matters. In a heavily negative nodule dataset, predicting “non-nodule” for everything produces a high percentage correct but misses every nodule. Classification precision asks a narrower question: when the model predicts a nodule, how often is it right? section 14.0 section 14.3.2
Precision is true positives divided by all positive predictions: true positives / (true positives + false positives). A true positive is a correctly predicted nodule; a false positive is a non-nodule incorrectly predicted as a nodule. High precision means few false alarms among the model’s positive predictions. section 14.3.2 section 14.3.3
A cautious classifier
Imagine a guard dog that barks only when nearly certain it has spotted a burglar. Almost every bark is useful, so its precision is high. But it may let many burglars pass unnoticed. Likewise, raising a classifier’s threshold can improve precision while hurting recall; precision alone cannot establish whether the classifier is useful. section 14.3.2 section 14.3.4
Calculate and inspect it during training
Count correct positive predictions as truePos_count, and negative samples incorrectly predicted positive as falsePos_count. Then compute precision = truePos_count / np.float32(truePos_count + falsePos_count) and log it each epoch alongside recall and F1 score. Check the denominator: when the model predicts nothing positive, both counts can be zero and precision becomes nan. A high overall accuracy can hide that failure. section 14.3.3 section 14.3.5
Recall measures a different failure: missed positives. The F1 score combines precision and recall because changing the classification threshold can make either one look good at the other’s expense. section 14.3.4
