Figure 2.

Experimental design. For each patient, the acquired CXRs formed a series of sequential pairs. For each pair (example shown for CXR2 and CXR3), an outcome was defined by judging if the radiological evolution of second CXR of the pair was worse, stable or improved compared to the first. We then tested whether the difference ($\mathrm{\Delta }$ CXR) in radiological findings probabilities would be statistically different between outcome categories; and secondly if deep learning features from the first CXR of the pair would predict radiological evolution.