Min Jung Kim

Concept accuracy can be misleading, as relying solely on prediction performance may mask critical deficiencies in the physical scale of intermediate variables in clinical models.

Spatially faithful models can still overlook crucial temporal information, raising concerns about the reliability of current attribution methods in medical diagnostics.

Forget painstakingly labeling mammograms: this method uses synthetic data and generative refinement to segment microcalcifications with surprising accuracy, even across different datasets.