Chest radiograph classification is useful to support

research on automated pneumonia screening but apparent ben-

efits should be analysed in consideration of the composition of

the databases and specificities and sensitivities. In this work,

a three-block CNN trained from scratch is compared with a

ResNet18 classifier which is initialised from the ImageNet, and

examples of predictions explained using gradient-weighted class

activation mapping (Grad-CAM) is shown. Recovered processes

RGB-format images with a resolution of 224 × 224, and evaluates

the prediction result of normal vs. pneumonia image in a

two-class image folder, with cross-entropy optimization. The

saved notebook contains a 2,831-image loaded training set and

a 624-image evaluation confusion matrix. The first reported

accuracies are 75.64% for the CNN and 85.74% for ResNet18.

The confusion matrix saved with ResNet18 is {388 True Positives,

87 False Positives, 147 True Negatives, 2 False Negatives} with

the following scores: sensitivity = 99.49%, specificity = 62.82%,

precision = 81.68%, and F1 score = 89.71%. The qualitative

Grad-CAM overlays show model focus of activation but are

not medically validated localisation. The evaluation folder was

used as a validation time and again, there is an incomplete run

provenance, and the method is not clinically validated. These

findings encourage an applied example of transfer learning and

explainability in qualitative terms and offer further impetus to

an independent evaluation.