Refining YOLOv8 for Full Field Digital Mammograms: Improving Small Object Detection through Resolution-Preserving Patched Inference
Itana Bulatović, Guangming Wang, Hesheng Wang · 2025
This research presents a novel approach to enhance mammography assessment and improve early breast cancer detection, with focus on detecting tiny objects such as microcalcifications. Building on the state-of-the-art one-stage detector YOLOv8, we propose a specialized model architecture that incorporates a high-resolution detection block with a Convolutional Block Attention Module (CBAM). These additions enable the model to better capture and distinguish fine details critical for identifying small structures in mammograms. Furthermore, to address the challenge of resolution loss typically associated with compressing large full-field digital mammography images, we introduce a Patched Inference (PI) technique. This approach maintains full image resolution and preserves feature quality, enhancing Small Object Detection (SOD) accuracy for microcalcifications and small masses. To mitigate overlapping predictions that can arise from patched inference, we propose a post-processing algorithm called Maximum Box Fusion (MBF), which fuses overlapping detection boxes to improve prediction accuracy, especially for large objects. The proposed workflow is rigorously evaluated on two recent mammography datasets, VinDr-Mammo and INbreast, for microcalcification and mass detection, across various object sizes.