Fig. 1

A detailed description of mosaic formation and soft labeling process. a A set of four images consisting of a mosaic image. One image is selected for each class and the last one is additionally and randomly selected from all classes (in this case, the noncancerous lesion). b A Grad-CAM-based ROI extraction proposes a representative region of each image. A class activation map of each image is extracted from the trained model without mosaic dataset. c The mosaic generator randomly chooses a point within a square to form a 2 × 2 grid so that one segment should be, at least, larger than half of the synthesized image. A mosaic image is synthesized with four cropped and rescaled images from each Grad-CAM-based ROI. The position of each image is randomly assigned each time a synthesized image is created. Also, the oversampling rate is adjustable, allowing for larger areas to be allocated to user-defined classes. (in this case, A: OPMD, B: noncancerous lesion, C: cancer, D: noncancerous lesion) d The soft label is calculated with the areas of each class within the grid. The model is trained to accurately predict the value of each component in the soft label vector