Enhancing convolutional neural network model with spectral features for the identification of cervical dysplasia
Sabeena Karim Kutty, Gopakumar Chandrasekhara Menon · International Journal of Imaging Systems and Technology · 2022
Abstract Cervical cell classification plays a key role in the computer‐based screening and diagnosis. This article focuses on cell classification and grading to differentiate the stages of cervical dysplasia. The proposed framework uses a unification of wavelet transform and convolutional neural network (CNN) for segregating spectral and spatial features from papanicolaou stained (pap) smear images. A correlation‐based feature selection is adopted to find relevant features from the CNN model. Random Forest classifier is then incorporated to identify cervical dysplasia from the extracted features. The proposed framework is verified on three publicly available datasets. Using the first dataset, the method obtained an average accuracy of 98.12% and area under curve (AUC) of 0.999. In the second (Sipakmed) and third (Herlev) dataset that consists of cell images of both normal and dysplastic condition, the method achieved an accuracy of 97.01%, 82.60% and AUC of 0.999 and 0.907, respectively. The observations of the study imply that spectral features with feature selection along with a proven machine learning model improves the detection rate of cervical dysplasia.