Statistical distortion of supervised learning predictions in optical microscopy induced by image compression.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Optical microscopy generates large datasets, necessitating data compression for automated analysis using supervised learning (SL).
• Assessing the impact of image compression on SL model reliability is critical, particularly for clinical applications.

Purpose of the Study:

• To quantify statistical distortions in SL predictions caused by image compression.
• To evaluate the reliability of SL models when applied to compressed microscopy data.

Main Methods:

• Compared SL predictions on compressed data with raw predictive uncertainty derived from sensor noise statistics.
• Quantified alterations in cell segmentation parameters after pixel depth reduction and JPEG compression.

Main Results:

• Pixel depth reduction (16-to-8 bits) and 10:1 JPEG compression altered cell segmentation predictions by up to 15% and over 10 standard deviations.
• Higher JPEG compression ratios led to significantly larger distortions.
• A metrologically accurate compression algorithm yielded prediction spreads comparable to raw noise.

Conclusions:

• Image compression introduces significant statistical distortions affecting SL model performance in optical microscopy.
• The developed method establishes a lower bound for predictive uncertainty in SL tasks.
• This approach can generalize to assess distortions from various processing pipelines in AI-assisted fields.