Twenty Scans in a Thousand

Accuracy ranks them backwards: the always-healthy model scores $98.0\%$ against this model's $97.4\%$, yet catches none of the $20$ cases while this one catches $12$.

The always-healthy model is genuinely the better of the two here, since $98.0\%$ beats $97.4\%$ and it never raises a single false alarm on any of the $1000$ scans.

The two are close enough to be treated as equivalent, since a gap of $0.6$ percentage points is well within the noise you would expect on a set of only $1000$ scans.

The lower accuracy is a symptom of overfitting, and training for more epochs on the $980$ healthy scans would be enough to close the $0.6$ point gap.