Embodiments of the disclosure are directed to segmenting a digital image of biological tissue into biological units, such as cells. A first weak or data driven segmentation is generated using image data representing the digital image to segment the digital image into a first set of biological units. Applying a geometric model, each unit in the first set of biological units is ranked based on a similarity in shape and scale between the unit and one or more other units in the image. A subset of units from the first set of biological units is selected based on the rank of each biological unit relative to a predetermined threshold rank. A second weak or data driven segmentation may then be generated using image data including the subset of biological units to segment that portion of the digital image into a second set of biological units.