Semi-Supervised Cell Detection with Reliable Pseudo-Labels.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Pathological images are crucial for cancer diagnosis and prognosis.
• Manual analysis of these images is time-consuming and labor-intensive.
• Obtaining fully annotated pathological image datasets is challenging due to data scarcity.

Purpose of the Study:

• To address the issue of insufficient labeled data in pathological image analysis.
• To develop methods for generating accurate pseudo-labels from unlabeled pathological images.
• To improve the performance of cell detection models using unlabeled data.

Main Methods:

• Proposed two pseudo-labeling methods: adaptive threshold and cell count.
• Utilized data distillation with an attention mechanism for model retraining.
• Developed a multi-task learning model for simultaneous cell detection and counting.

Main Results:

• Improved cell detection by 9%-13% when incorporating unlabeled data compared to using only labeled data.
• Demonstrated the effectiveness of adaptive threshold and cell count pseudo-labeling.
• Showcased the applicability of the proposed methods across three different datasets.

Conclusions:

• The proposed pseudo-labeling and multi-task learning approaches effectively leverage unlabeled pathological images.
• These methods significantly enhance the accuracy and efficiency of cancer cell detection.
• The developed techniques offer a practical solution for improving pathological image analysis in clinical settings.