Towards Generalizability and Robustness in Biological Object Detection in Electron Microscopy Images

Abstract

A major barrier in the adoption of machine learning models is thelack of robustness and generalizability, when the data that the modelsare applied to have different characteristics from those they aretrained on. This is especially the case in the biomedical field, wherelimited data size, data volatility and data shifts are common. Withoutproper tuning and data normalization, deploying machine learningmodels in the presence data shift leads to reduced or misleading performance.This study explores techniques to enhance model generalizabilitythrough iterative adjustments, specifically, with different normalizationand augmentation techniques, for the task of object detecion in electronmicroscopy images. We found that models trained with GroupNormalization or texture data augmentation outperform other normalizationtechniques and classical data augmentation, enabling themto learn more generalized features. These improvements persist evenwhen models are trained and tested on disjoint datasets acquiredthrough diverse data acquisition protocols. Results hold true for bothtransformer- and convolution-based detection architectures. The experimentsshow an impressive 29% boost in average precision, indicatingsignificant enhancements in the model’s generalizibality. This underscoresthe models’ capacity to effectively adapt to diverse datasetsand demonstrates their increased resilience in real-world applications.

Keywords: Data normalization, data shift, group normalization, dataaugmentation, robustness, generalization, electron microscopy, object detection