Key Achievements

Model Adaptation

Started with an existing Keras-based Mask R-CNN model designed for detecting six common types of cell nuclei in NSCLC ( Non-small cell lung cancers) pathology images. Through transfer learning, the model was fine-tuned to detect and classify seven common types of cell nuclei in breast cancer pathology images, achieving simultaneous segmentation and classification.

Data Labeling Challenge

Faced with a dataset where approximately 20% of the samples had incomplete labels—missing either the class labels (unlabeled) or the pixel-level mask labels—direct training was not feasible without addressing the issue.

Loss Function Redesign

Successfully redesigned the Mask R-CNN’s loss function to selectively ignore samples with missing labels during the computation of losses. Specifically, samples lacking classification labels were excluded from classification loss calculations, and those missing mask labels were not considered in mask loss calculations.

Performance

The modified Mask R-CNN model, now based on the PyTorch framework, demonstrated a detection rate of 82.5% and a 7-class classification accuracy of 82.0% for breast cancer cell nuclei in H&E stained pathology images.

Why Mask R-CNN?

1. Precision in Instance Segmentation: Mask R-CNN is renowned for its high precision and accuracy in instance segmentation tasks, crucial for accurately locating and segmenting cell nuclei in tumor pathology.
2. Detailed Object Masks: Beyond detecting object boundaries, Mask R-CNN generates detailed instance masks, providing valuable insights into the shape and structure of tumor cell nuclei.
3. Flexibility and Customizability: The PyTorch-based implementation of Mask R-CNN offers the flexibility needed to tailor the model to specific requirements, making it ideal for handling tumor cell nuclei of varying shapes, sizes, and densities.
4. Community Support: Mask R-CNN benefits from extensive community support, including a wealth of resources, documentation, and pre-trained models, facilitating the development process and accelerating model training and optimization.

Conclusion

This project showcases the potential of advanced machine learning models like Mask R-CNN in enhancing the detection and classification of cell nuclei in pathology images, contributing to more accurate and efficient diagnostic processes in oncology.