Neuroblastoma differentiation type diagnosis algorithm based on Dense-U-Net using whole slide images

• Published in: Signal, image and video processing Vol. 18; no. 5; pp. 4627 - 4635
• Main Authors: Wan, Zhenzhen, Liu, Yuwei, Liu, Fang, Shi, Ning, Zhang, Nan, Liu, Xiuling
• Format: Journal Article
• Language: English
• Published: London Springer London 01.07.2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Computer Imaging; Computer Science; Datasets; Deep learning; Diagnosis; Differentiation; Encoders-Decoders; Heterogeneity; High resolution; Histopathology; Hospitals; Image Processing and Computer Vision; Image resolution; Medical diagnosis; Morphology; Multimedia Information Systems; Neural networks; Neuroblastoma; Original Paper; Pancreatic cancer; Pathology; Pattern Recognition and Graphics; Semantics; Signal,Image and Speech Processing; Tumors; Vision; Deep learning; Whole slide image; Neuroblastoma; Differentiation
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-024-03100-9

Neuroblastoma is a type of peripheral neuroblastic tumor and is a common malignant solid tumor in children with significant biological heterogeneity and rapid development. Determining the type of differentiation is important in predicting the prognosis of neuroblastoma and making early judgments on postoperative treatment plans. The whole slide images of neuroblastoma offer high resolution that greatly facilitates clinical reading. However, the high resolution of these images makes computer-aided diagnosis quite challenging. To address the challenge, we propose in this paper a network model based on the traditional encoder-decoder structure of U-Net. The model uses DenseNet as the encoder to extract features of image, initializes network weights by transfer learning, and builds the decoder based on the up-sampling module and dense block to extract different types of cells. K-means algorithm is used to cluster and count the poorly differentiated cells and differentiated cells, identifying the histopathological types and the prognostic effects. Our experimental results demonstrate that our model performs excellently, having achieved 97.55% of segmentation accuracy, 92.10% and 89.02% counting accuracy for poor differentiated cells and differentiated cells, respectively.