Rotation Invariance and Extensive Data Augmentation: A Strategy for the MItosis DOmain Generalization (MIDOG) Challenge

Abstract

Automated detection of mitotic figures in histopathology images is a challenging task: here, we present the different steps that describe the strategy we applied to participate in the MIDOG 2021 competition. The purpose of the competition was to evaluate the generalization of solutions to images acquired with unseen target scanners (hidden for the participants) under the constraint of using training data from a limited set of four independent source scanners. Given this goal and constraints, we joined the challenge by proposing a straight-forward solution based on a combination of state-of-the-art deep learning methods with the aim of yielding robustness to possible scanner-related distributional shifts at inference time. Our solution combines methods that were previously shown to be efficient for mitosis detection: hard negative mining, extensive data augmentation, rotation-invariant convolutional networks.

We trained five models with different splits of the provided dataset. The subsequent classifiers produced F₁-score with a mean and standard deviation of $0.747 {\pm } 0.032$ on the test splits. The resulting ensemble constitutes our candidate algorithm: its automated evaluation on the preliminary test set of the challenge returned a F₁-score of 0.6828.