DEEP NEURAL NETWORK FOR 3D PARTICLE DETECTION IN 3D FLUORESCENCE MICROSCOPY IMAGES VIA DENSITY MAP REGRESSION

Automatic detection of particles in fluorescence microscopy images is crucial to analyze cellular processes. We introduce a novel deep learning method for 3D fluorescent particle detection. Instead of pixel-wise binary classification or direct coordinate regression, we perform image-to-image mapping based on regressing a density map. Detections close to particles are rewarded in the network training, and highly nonlinear direct prediction of point coordinates is avoided. To focus on particles in comparison to background image points, we suggest using the adaptive wing loss. We also employ a weighted loss map to cope with the very strong imbalance between particle and background image points for 3D images. We evaluated our approach using 3D images of the Particle Tracking Challenge and real 3D fluorescence microscopy images of chromatin structures and interneurons. It turned out that our approach generally outperforms previous methods.