YOLO-TL: A Tiny Object Segmentation Framework for Low Quality Medical Images

Accurate segmentation of cancer cell nuclei in medical images is fundamental for clinicians to diagnose and evaluate tumors, enabling scientists to explore genetic variants and epigenetic changes in cancer cells and reveal the complex mechanisms of tumorigenesis. Fine cell nucleus segmentation not only improves the accuracy of cancer diagnosis, but is also an important tool for personalized treatment and advances in cancer research. Currently, the representatives of image segmentation models in the field of medical image analysis, various improved versions of U-Net, exhibit suboptimal performance on certain datasets. The challenges arise from the non-ideal shooting conditions prevalent in medical imaging, characterized by low-light environments and low-pixel slices. Compounding the issue is the diminutive size of cell nuclei relative to the overall image, coupled with their potentially abundant and densely distributed nature within the frame. Based on these difficult thinking on, a method is used SPD convolution to construct a new multi-feature fusion layer, and combined with the SimAM attention mechanism to improve the current YOLOv8 model. YOLO-TL achieves a Dice score of 0.852 and a Jaccard coefficient of 0.7804, where the Dice score is at least 2% points or more higher than a host of improved U-Net-based models.