Dynamic Incremental Learning in Medical Imaging

The foundational models such as YOLO, Faster R-CNN, and SSD have demonstrated substantial potential in computer vision, particularly in object detection. However, their application in medical imaging faces significant challenges due to the need for fine-grained recognition of small lesions and the continuous emergence of new disease types. Directly extending these models to medical image object detection is difficult, as they often struggle with generalizing to unseen radiological findings. Additionally, incremental learning in this domain is hampered by issues such as catastrophic forgetting and background shift. Our study proposes an incremental learning-based approach to enhance the performance and stability of medical image object detection models. This method addresses the unique challenges of medical imaging, including small lesion detection and low visual contrast, by combining dynamic weighted loss, augmented pseudo labels, and confidence score distillation. Incremental learning allows the model to adapt to new data without forgetting previously learned knowledge, which is essential given the limited size and evolving nature of medical imaging datasets. The main contributions of this study include the development of a dynamic weighted loss and network tailored for medical images, which improves detection accuracy. We also generate reliable pseudo-labels and apply data augmentation to address background shift and sample imbalance. Furthermore, we introduce an object confidence knowledge distillation method to enhance detection stability. Our approach demonstrates significant improvements in detection performance, showcasing its potential for practical clinical application and its implications for future medical image analysis and diagnosis.