(P 244) A DEEP FEW-SHOT META-LEARNING STRATEGY FOR AUTONOMOUS CLASSIFICATION OF BONE METASTASIS IN RADIONUCLIDE BONE SCANS

Objective: Automatic classification of bone metastasis using only radionuclide scans will facilitate communication among radiologists and surgeons, offering a standardized approach to cope with artificial intelligence (AI)-powered medical image analysis. The objective of this work is to present a novel deep few-shot meta-learning strategy for the autonomous classification of bone metastasis in radionuclide bone scans. The primary aim is to address the challenges posed by limited labeled data by leveraging the power of meta-learning, enabling the model to quickly adapt to new classes with only a few sample images. Our approach combines deep neural networks and meta-learning techniques to achieve high accuracy and robustness in classifying bone metastasis, even with scarce labeled data. This work can facilitate early and accurate diagnosis, as well as aid in personalized treatment planning for cancer patients where the gold-standard classification required for training is lacking or expensive to achieve.

Methods: Deep learning methods generally require large number of images to learn a feature representation from medical images. We implemented a deep few-shot meta-learning pipeline using prototypical networks to perform both 5-way 10-shot and 5-way 25-shot classification. The goal is to be able to quickly adapt and automatically predict bone metastasis in unseen classes throughout different regions of the body. Various meta-learning tasks were created for training, validation, and testing with no overlapping regions in each split with 18 different classes in training, and 9 different classes in both validation and testing, with all across chest, head, knee, pelvis, shoulder, and vertebra. All images and the class labels were collected and constructed from the open-access BS-80K dataset (DOI: 10.1016/j.compbiomed.2022.106221).

Results: For 10-shot 5-way classification and 25-shot 5-way classification, 500 test tasks were randomly sampled. The average accuracy (AUC) for 10-shot 5-way classification was 0.928 +/- 0.085, while the accuracy for 25-shot 5-way classification was 0.932 +/- 0.081.

Conclusion: Based on the experimental results, we conclude that our deep few-shot meta-learning strategy exhibits promising performance in the autonomous classification of bone metastasis in radionuclide bone scans, offering potentials of quickly classifying new unseen images using only a few labeled images. This is a viable approach in settings where the gold-standard images required for training is lacking or expensive to achieve in a clinical setting. While our work in the future will further focus on the implementation and uptake of this approach, we will aim to incorporate multiple clinical data into this pipeline to facilitate personalized treatment planning for cancer patients.