A Systematic Comparison of Task Adaptation Techniques for Digital Histopathology

Area of Science:

• Computational histopathology
• Artificial intelligence in medicine
• Computer vision

Background:

• Insufficient image annotation in histopathology necessitates fine-tuning pre-trained neural networks.
• Vanilla fine-tuning is common, but newer computer vision algorithms may offer improved accuracy.
• Existing research shows benefits for medical AI (radiology), but empirical evidence in histopathology is lacking.

Purpose of the Study:

• To systematically compare nine AI task adaptation techniques against vanilla fine-tuning in histopathology.
• To evaluate the accuracy improvements of advanced methods on diverse histopathological classification tasks and datasets.
• To investigate the impact of training set size on the performance of selected techniques.

Main Methods:

• Utilized the ConvNeXt architecture for comparison.
• Evaluated nine task adaptation techniques: DELTA, L2-SP, MARS-PGM, Bi-Tuning, BSS, MultiTune, SpotTune, Co-Tuning, and vanilla fine-tuning.
• Conducted systematic comparisons across five histopathological classification tasks and eight datasets, including external testing and statistical validation.

Main Results:

• Five advanced task adaptation techniques demonstrated significant relative accuracy improvements over vanilla fine-tuning.
• Technique suitability varied depending on the specific histopathological classification task.
• Co-Tuning showed notable performance gains (e.g., P(≫) = 0.942, d = 2.623).
• Analysis of training set size revealed varying impacts on accuracy for methods like Co-Tuning (e.g., P(≫) = 0.951, γ = 0.748).

Conclusions:

• Advanced task adaptation techniques offer significant potential for improving AI accuracy in histopathology.
• The effectiveness of these techniques is influenced by factors like the classification task and training data availability.
• Tailoring AI methods to specific histopathological challenges is crucial for optimal performance.