A Classification Method for the Cellular Images Based on Active Learning and Cross-Modal Transfer Learning.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Accurate classification of human epithelial type 2 (HEp-2) cells is crucial for diagnosing autoimmune diseases using computer-aided diagnosis (CAD) systems.
• Current supervised learning methods require extensive manual annotation, posing a significant challenge for developing robust HEp-2 cell classifiers.

Purpose of the Study:

• To develop an active learning strategy to minimize the data annotation burden in HEp-2 cell classification.
• To improve the efficiency of CAD systems by reducing the need for large, manually labeled datasets.

Main Methods:

• A novel approach combining cross-modal transfer learning with parallel deep residual networks.
• Initial training of parallel networks using a small, annotated dataset with diverse wavelet coefficients as input.
• Application of active learning on a larger, unannotated dataset to intelligently select images for annotation.

Main Results:

• Demonstrated that active learning significantly reduces the number of required annotated examples.
• Achieved comparable discrimination performance to state-of-the-art methods with substantially less labeled data.
• Validated the effectiveness of combining active learning with efficient transfer learning techniques.

Conclusions:

• Active learning, integrated with cross-modal transfer learning and parallel deep residual networks, offers an efficient solution for HEp-2 cell classification.
• This approach simplifies the development of CAD systems by alleviating the laborious image labeling process.
• The proposed method maintains high classification performance while drastically reducing annotation efforts.