One Exposure, Several Readings: How Firm Is an AI Verdict? Mammography and Other Machines That Rely on AI

An AI program reads a mammogram and gives a score. Today the reader sees one score from one picture, and does not know how firm that score is. This note describes a check that can sit on top of any such program, on the hardware of the machine or connected to it. The woman is X-rayed once, and the raw data of that exposure is kept exactly as it is. A small mathematical algorithm then moves the settings of the set-up a little and reads the score at each. There are two ways. In Way 1, the settings of the image-processing module are moved, and the AI reads each version. In Way 2, the settings of the AI model itself are moved, and it reads the same processed picture at each. The number of versions is left to the computing power available. If the scores stay together, the score is firm. If they spread out, the score depends on a setting that should not matter. A computer-decided shunting action then follows. The computer decides without waiting for a person, and it moves the work aside instead of stopping it, while all other exams carry on. In Way 1, the exam is sorted into one of three states, Steady, Watch and Look closer, each with its own action. In Way 2, the computer moves to a new setting vector and reads again. If the scores still do not settle, the exam goes to the front of the radiologist's list. No second visit, no extra radiation and no change to the AI program are needed. The note explains how this differs from earlier work on mammography systems, test-time augmentation, Bayesian networks and Monte Carlo dropout, and deep ensembles. It then shows how the same idea applies to oral cancer, autism screening, Alzheimer's brain scans, ECG, MRI, CT scans, chest X-rays, DNA fingerprinting, forensic analysis, border surveillance and battery management systems. Keywords: mammography, AI uncertainty, firmness of a verdict, shunting action, medical imaging, forensic analysis, surveillance, battery management