M2IoU: A Min-Max Distance-based Loss Function for Bounding Box Regression in Medical Imaging

Computer vision applications such as object detection have increased manifolds in the medical domain for diagnosis and treatment purposes. Generally, object detection models such as YOLO (You Only Look Once) involve identifying the correct bounding box and classifying the objects inside the bounding box. However, medical imaging object detection is a challenging endeavor, requiring models that are both efficient and extremely accurate in the face of limited data and expensive annotations. In this paper, we propose Min-Max IoU (M2IoU) loss function by introducing a new min-max-based penalty term in the loss equation, between the predicted box and the ground truth coordinates. We further compare the results of several loss functions on the YOLOv8 model trained on multiple medical datasets and demonstrate that the M2IoU loss function leads to faster learning and outperforms other existing loss functions like CIoU and GIoU.