HscoreNet: A Deep network for estrogen and progesterone scoring using breast IHC images. (June 2020)

• Record Type: Journal Article
• Authors: Saha, Monjoy; Arun, Indu; Ahmed, Rosina; Chatterjee, Sanjoy; Chakraborty, Chandan
• Is Part Of: Pattern recognition. Volume 102(2020:Jun.)
• Publication Date: 2020-06
• Subjects: Breast -- Estrogen -- Progesterone -- Encoder -- Decoder; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107200
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Deep learning (DL) architecture has been proposed for automated scoring of Estrogen and Progesterone using breast cancer IHC Images. The proposed DL architecture consisted of encoder, decoder, and scoring layer. The concept of the scoring layer is new. Results show broad applicability and state-of-the-art performance. Graphical abstract: Abstract: Estrogen and progesterone receptors serve as an important predictive and prognostic biomarkers for breast cancer immunohistological analysis. For breast cancer prognosis, pathologists manually compute the score based on the visual expression and the number of immunopositive and immunonegative nuclei. This manual scoring technique is time-consuming, cumbersome, expensive, error-prone, and susceptible to intra- and interobserver ambiguities. To solve these issues, we proposed a deep neural network (i.e., HscoreNet ), which consists of three parts, i.e., encoder, decoder, and scoring layer. A total of 600 (300 ER and 300 PR) regions of interest at 40 × magnification from 100 histologically confirmed slides were used in this study. The size of each region of interest was 2048 × 1536 pixels (width × height). The encoder layer has been used to transform input pixels into a lower-dimensional representation, whereas the decoder reconstructs the output of the encoder through minimization of a cost function. The decoder generates an image that only contains immunopositive and immunonegative nuclei. The output of the decoder is fed to the input of the scoring layer. This layer computes the Histo-score or H-score based on the staining intensity, the color expression, and the number of immunopositive and immunonegative nuclei. Pathologists compute this score to subcategorize the cancer grades and to decide proper treatment procedures. Our proposed approach is affordable, accurate, and fast. We achieved excellent performance, with 95.87% precision and 94.53% classification accuracy. Our proposed approach streamlines the human error-prone and time-consuming process. This methodology can also be used for other types of histology and immunohistology image segmentation and scoring.