A meta-learning based method for few-shot pneumonia identification using chest X-ray images

Pneumonia is a prevalent respiratory disease and conventional manual diagnosis methods are prone to misdiagnosis. Recently, deep learning based computer-aided diagnosis systems have played an important role in pneumonia identification. However, these methods face challenges in scenarios where there are limited samples available, as they require a substantial number of labeled samples for training to achieve optimal performance. In practical scenarios for diagnosing pneumonia, the availability of labeled samples is scant due to the high cost of labeling, the imperative to protect patient privacy, and the low incidence rates of rare diseases. To address this challenge, we propose a novel meta-learning method that incorporates disease hierarchical relationships and similarities as additional domain knowledge to guide model learning thereby enhancing diagnostic accuracy. This method formulates a disease hierarchy tree based on the International Classification of Diseases (ICD-11) and designs a hierarchical classification module to classify diseases at multiple granularities based on the hierarchy tree. After establishing an effective feature extractor, a related task module is proposed to compute disease similarities and construct related tasks from similar diseases. These tasks are utilized for fine-tuning the model to maximize the transfer of diagnostic knowledge. Experiments on the public dataset consisting of multiple types of pneumonia show that our proposed method achieved the highest classification accuracy in both 1-shot and 5-shot settings.