Fig. 2: Schematic diagram of the rare event detection (RED) pipeline.
From: Unsupervised detection of rare events in liquid biopsy assays
In Step 1, one IF image is split into ≈ 2.5 million non-overlapping tiles. In Step 2, pairs of synthetically generated noisy tiles and their clean counterparts are used to train a denoising autoencoder (DAE). In Step 3, noisy tiles are used as input to the trained DAE and the difference between the de-noised and the original clean version of the tiles is used in combination with user-specified IF channel weights to evaluate the reconstruction error for each tile. Tiles with large values of the reconstruction error are identified and are deemed as being rare. In Step 4, an approach that assumes that true rare events are unlikely to be localized to a region within an IF image is used to eliminate artifacts from the rare tile cohort.
