KongNet tops three medical imaging challenges with multi-headed nuclei detection

KongNet is a novel multi-headed deep learning architecture designed for accurate nuclei detection and classification in histopathology images. Its design features a shared encoder that feeds into parallel, cell-type-specialized decoders, each jointly predicting nuclei centroids, segmentation masks, and contours. The model incorporates Spatial and Channel Squeeze-and-Excitation (SCSE) attention modules and a composite loss function to enhance feature extraction and multi-task learning. This specialized multi-decoder approach allows KongNet to handle diverse tissue and stain types effectively, addressing a key challenge in computational pathology.

The model's performance was validated across three major Grand Challenges. In the MONKEY Challenge, KongNet achieved first place on track 1 and second place on track 2. Its lightweight variant, KongNet-Det, secured first place in the 2025 MIDOG Challenge. Additionally, a version pre-trained on MONKEY and fine-tuned on the PUMA dataset ranked among the top three in the PUMA Challenge without further optimization. KongNet also set new state-of-the-art results on the publicly available PanNuke and CoNIC datasets. The researchers have publicly released pre-trained model weights and inference code to support further research in medical image analysis.