A Robust Training Method for Pathological Cellular Detector via Spatial Loss Calibration.

Area of Science:

• Medical Imaging
• Computational Pathology
• Machine Learning

Background:

• Accurate diagnosis from pathological images requires detecting all positive cells.
• Cellular datasets often have sparse annotations due to annotation challenges.
• Training detectors on sparse data can lead to miscalculated losses and reduced performance.

Purpose of the Study:

• To develop an efficient and reliable method for training cellular detectors on sparsely annotated datasets.
• To address the issue of miscalculated losses in detectors trained on sparse annotations.

Main Methods:

• Proposed a novel training method utilizing regression boxes' spatial information.
• Implemented loss calibration to mitigate miscalculated losses.
• Evaluated performance on datasets with varying degrees of sparse annotations.

Main Results:

• The proposed method significantly boosts detector performance on sparsely annotated datasets.
• Performance remained robust even with up to 90% missing annotations.
• Identified a strong correlation between middle detector layers and generalization performance.

Conclusions:

• The developed method offers an effective solution for training cellular detectors with sparse annotations.
• The findings provide insights into the relationship between network layers and generalization.
• Suggests future research directions for layer-specific constraint rule design using gradient analysis for precise model training.