Impact of imperfect annotations on CNN training and performance for instance segmentation and classification in...

Area of Science:

• Digital pathology
• Computational biology
• Machine learning in healthcare

Background:

• Accurate segmentation and classification of cell nuclei are vital for diagnosing diseases from histopathology images.
• Deep learning models require large, high-quality annotated datasets, which are difficult and time-consuming to create.

Purpose of the Study:

• To investigate the impact of noisy annotations on Convolutional Neural Network (CNN) model performance for nuclei detection, segmentation, and classification.
• To determine optimal training strategies, specifically the number of training epochs, to mitigate overfitting to noisy labels.

Main Methods:

• Trained a state-of-the-art CNN model on histopathology images with varying degrees of noisy annotations.
• Evaluated model performance using a small, meticulously annotated validation set.
• Investigated the effect of pre-training on model robustness.

Main Results:

• Noisy annotations significantly degrade CNN model performance for nuclei analysis.
• Utilizing a small, clean validation set effectively prevents overfitting to annotation noise.
• Pre-training the CNN model demonstrated a beneficial role in improving robustness and overall performance.

Conclusions:

• A small, correctly annotated validation set is crucial for training robust deep learning models in digital pathology.
• Pre-training enhances model performance and resilience against noisy annotations in histopathology image analysis.