Lesion-Harvester: Iteratively Mining Unlabeled Lesions and Hard-Negative Examples at Scale

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Acquiring large annotated medical image datasets for machine learning is costly and time-consuming.
• Existing datasets like DeepLesion often have incomplete or noisy labels, hindering model training.
• Effective methods for automatically harvesting or discovering missing annotations are crucial.

Purpose of the Study:

• To develop and validate Lesion-Harvester, a system for high-precision annotation harvesting from lesion datasets.
• To improve the completeness and quality of medical image annotations for machine learning.
• To enhance the performance of downstream machine learning models through augmented datasets.

Main Methods:

• A chained system combining a sensitive lesion proposal generator (LPG) and a selective lesion proposal classifier (LPC).
• Iterative finetuning of the LPG using a novel hard negative suppression loss.
• Development of a 3D contextual LPG and a global-local multi-view LPC for optimized performance.
• Introduction of a pseudo 3D Intersection over Union (IoU) evaluation metric.

Main Results:

• Lesion-Harvester discovered an additional 9,805 lesions in the DeepLesion dataset with 90% precision.
• Augmenting DeepLesion annotations with harvested lesions improved state-of-the-art detectors' average precision by 7-10%.
• Public release of harvested lesions and a fully annotated DeepLesion test set.

Conclusions:

• Lesion-Harvester effectively addresses the challenge of incomplete annotations in medical imaging datasets.
• The system significantly enhances the utility of existing datasets for machine learning model development.
• The released data and methods facilitate further research in automated medical image annotation and analysis.