Features in Backgrounds of Microscopy Images Introduce Biases in Machine Learning Analyses.

Area of Science:

• Pharmaceutical Science
• Analytical Chemistry
• Biotechnology

Background:

• Subvisible particles are critical in therapeutic protein formulations.
• Flow imaging microscopy (FIM) and backgrounded membrane imaging (BMI) are key techniques for particle analysis.
• Both FIM and BMI capture digital images of particles for characterization.

Purpose of the Study:

• To compare the performance of convolutional neural networks (CNNs) for classifying subvisible particles using FIM and BMI data.
• To investigate the influence of image background on CNN classification accuracy for both techniques.
• To assess the robustness of CNN models trained on FIM versus BMI images.

Main Methods:

• Convolutional neural networks (CNNs) were trained to classify particle images from FIM and BMI.
• Attribution analyses were performed to understand feature importance in CNN predictions.
• Image segmentation was used to reduce background influence on CNN performance.

Main Results:

• CNNs trained on BMI initially showed higher accuracy than those trained on FIM.
• Attribution analysis revealed BMI-CNNs relied on membrane background features, while FIM-CNNs focused on particle features.
• Image segmentation decreased BMI-CNN accuracy significantly but minimally impacted FIM-CNN accuracy.

Conclusions:

• The superior accuracy of BMI-trained CNNs was an artifact of background features, not particle characteristics.
• Robustness checks using attribution methods are crucial for validating machine learning models in particle analysis.
• Careful consideration of image acquisition techniques and potential artifacts is essential for reliable subvisible particle characterization.