Multi-feature fusion of deep networks for mitosis segmentation in histological images

Mitotic cell detection in pathological images is significant for predicting the malignancy of tumors and the intelligent segmentation of these cells. Overcoming human error generated by pathologists in reading the images while enabling fast detection through high computing power remains a very challenging task. In this study, we proposed a method that fuses handcrafted features and deep features to segment mitotic cells in whole-slide images. The handcrafted feature extraction strategy was based on four measure indices of the Gray Level Co-occurrence Matrix. The deep feature extraction strategy was based on natural image knowledge transfer. Finally, the two strategies were fused to classify and distinguish the image pixels for the segmentation of mitotic cells. We used the AMIDA13 dataset and the pathological images collected by the Department of Pathology of Anhui No. 2 Provincial People's Hospital as the experimental dataset. We compared the Areas Under Curve (AUC) of Receiver Operating Characteristic obtained through the handcrafted feature model, the improved deep feature model with knowledge transfer, the classic U-NET model, and the proposed multi-feature fusion model. The results showed that the AUC values of our proposed method had 0.07 and 0.05 improved to classic U-NET model on test dataset and validation dataset respectively, while achieved the best segmentation performance and detected most of true-positive cells, representing a breakthrough for clinical application. The experiments also indicated that the staining uniformity of pathological tissue impacted the model performance.