Deep Learning Models to Predict Fatal Pneumonia Using Chest X-Ray Images

Figure 4

Overview of the deployed deep learning model viewer by Google Cloud AutoML Vision and Kaplan–Meier plots for time to death from the diagnosis of pneumonia in the external validation test dataset. (a) A true-positive CXR image: the deep learning model accurately predicted it as a fatal case with a score of 0.64, and the actual prognosis was fatal. (b) A false-positive CXR image: the deep learning model predicted it as a fatal case with a score of 0.60, and the actual prognosis was nonfatal. (c) A false-negative CXR image: the deep learning model predicted it as a nonfatal case with a score of 0.75, and the actual prognosis was fatal. (d) A true-negative CXR image: the deep learning model predicted it as a nonfatal case with a score of 0.99, and the actual prognosis was nonfatal. (e) The deep learning model by Google Cloud AutoML Vision predicted nonfatal or fatal pneumonia using CXR images (a survival rate of 0.34 for patients with fatal pneumonia predicted versus 0.75 for those with nonfatal pneumonia predicted, ). CXR, chest X-ray.

(a)

(b)

(c)

(d)

(e)