Abstract:
The distribution and appearance of nuclei are essential bio markers for the diagnosis and study of
cancer. Despite the importance of nuclear morphology accurate segmentation and classification
of nuclei instances is still one of the most challenging tasks due to the wide occurrence of
overlapping, cluttered nuclei having blurred boundaries. Existing methods particularly focus
on region proposal techniques and feature encoding frameworks, however often fail to precisely
identify instances. In this paper we propose a simple yet effective model that precisely recognizes
instance boundaries as well as caters to exhaustive class imbalance problems, thus yielding
accurate class information for each nucleus. We have utilized nuclei pixel positional information
i.e its distance from contours for accurate shape estimation along with an object probability
score for filtering true nuclei pixels from the background. We have also proposed a novel loss
function that draws the same nuclei instance pixels function pulls together for learning an
object-based clustering bandwidth thus reinforcing the jaccardian index of the nuclei instance.
The network comprises a lightweight attention-aware feature fusion-based architecture having
separate instance probability, shape radial estimator, and classification heads. A compound
classification loss function is used that minimizes loss by assigning weighted loss to each class
according to type occurrence frequency thus mitigating major class imbalance issues existing
in most publicly available nuclei datasets.
