Leveraging immuno-fluorescence data to reduce pathologist annotation requirements in lung tumor segmentation using...

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Accurate tumor segmentation in non-small cell lung cancer (NSCLC) histology images is crucial but hindered by the scarcity of high-quality ground truth annotations.
• Generating precise annotations from pathologists is time-consuming and expensive, posing a significant bottleneck for developing robust segmentation algorithms.

Purpose of the Study:

• To explore alternative methods for generating tumor labels to train segmentation models.
• To determine the minimum size of pathologist annotations required when using a large dataset of low-cost, low-accuracy pan-cytokeratin (panCK)-based annotations for pre-training.
• To evaluate the generalizability of a trained model across variations in H&E staining and imaging protocols.

Main Methods:

• An Attention U-Net architecture was employed for tumor segmentation.
• The model was initially pre-trained on a large dataset of panCK-based annotations (10,326 mm²).
• Subsequently, the model was fine-tuned using a smaller, high-accuracy dataset of pathologist annotations (246 mm²).
• Model performance was compared against foundation models and evaluated for generalizability across 112 samples from three centers with varying scanner types and protocols.

Main Results:

• The panCK pre-training followed by fine-tuning achieved a mean intersection over union (mIoU) of 82% (95% CI: 77–87%).
• This approach enabled a 70% reduction in the requirement for pathologist annotations without compromising segmentation performance.
• PanCK pre-training outperformed standard foundation models, demonstrating superior performance for NSCLC tumor segmentation.

Conclusions:

• Leveraging large, cost-effective panCK annotations for pre-training is a viable strategy to overcome the annotation bottleneck in NSCLC segmentation.
• This method significantly reduces reliance on expert pathologist annotations, accelerating the development of accurate and generalizable segmentation tools.
• The study demonstrates the feasibility of creating robust segmentation models that maintain consistent performance across diverse H&E imaging conditions and patient cohorts.