Reducing the Effect of Incomplete Annotations in Object Detection for Histopathology

Denys Kaliuzhnyi, Dmytro Fishman, Mikhail Papkov

Keywords: object detection, missing annotations, histopathology

TL;DR: We show that modern object detectors can efficiently learn cell object detection in histopathology under missing annotation constraints. We suggest a simple hyperparameter adjustment that significantly diminishes the effect of incomplete annotations.

Abstract: Training neural networks for object detection usually requires decent amounts of data to produce great results. Apart from the image variety, the number of annotated objects is a crucial factor for success. In histopathology, the average annotation density is very high, resulting in resource-consuming data preparation for neural network training. We explore the effect of incomplete annotations in object detection. We show that modern object detectors, such as YOLO-v5, can effectively learn from histopathology datasets that lack up to 90% of annotations. Additionally, we suggest an easy model tuning setup to reduce the impact of incomplete annotations and enhance learning capability overall. We publish our code at https://github.com/DenysKaliuzhnyi/yolov5.