A classifier can look accurate while missing every case that matters. In nodule classification, negative samples greatly outnumber positive ones, so predicting “non-nodule” for everything produces a high percentage correct but finds no nodules. Recall answers a more useful question: of the actual positive cases, how many did the classifier catch? section 14.0 section 14.3.1
Recall is true positives divided by all actual positives: true positives / (true positives + false negatives). A true positive is a nodule correctly labeled positive; a false negative is a nodule missed by the classifier. Recall is also called sensitivity. Raising recall means missing fewer positives. section 14.3.1 section 14.3.3
A guard dog that rarely misses a burglar
Imagine a guard dog that barks whenever an event might be a burglar. By setting a permissive classification threshold, it catches nearly every burglar: recall approaches 1.0. But it also barks at many harmless events. High recall alone does not mean the classifier is useful; precision tells a different part of the story. section 14.3.1 section 14.3.2
Listing 14.2 Calculating recall and precision from classification counts
recall = metrics_dict['pr/recall'] = \ truePos_count / np.float32(truePos_count + falseNeg_count) precision = metrics_dict['pr/precision'] = \ truePos_count / np.float32(truePos_count + falsePos_count)
The listing calculates recall from true-positive and false-negative counts, alongside the corresponding precision calculation. section 14.3.3
Track recall alongside precision, loss, and percentage correct each epoch. If a classifier predicts every sample as negative, its recall is zero even when overall accuracy looks excellent. If there are no actual positives, the recall formula has a zero denominator and does not yield a defined value; metric calculations need attention to such cases. F1 combines recall and precision so that a strong result on just one cannot disguise a weak result on the other. section 14.3.3 section 14.3.4 section 14.3.5
