With the advent of deep neural networks, application of machine learning in multidisciplinary problems enhanced many folds. Many unsolvable problems previously sought as complex to compute are now made solvable by deep neural network techniques. Problems like protein folding by Alpha-fold and Alpha-Go are prime examples.

In this study six well known convolutional neural networks are applied for the classification of uterine cervix cancer cases for both seven class and two class classification. A primary dataset was also created by collecting raw slide samples form the leading medical institutes. The machine learning techniques do require set of well-crafted feature values representing the ground truth. Many times, these features fail to represent the ground truth. The deep neural networks can extract all the relevant features itself and those extracted features are used for final classification. In this work the convolutional neural networks are used for extraction of features which are the used for training shallow neural networks. The shallow neural networks used are Levenberg Marquardt neural network, One Step Secant and Scaled Conjugate gradient descent. The results indicated that among the 6 convolutional neural networks the ResNet50 is best and among the three shallow neural network Levenberg Marquardt is best for both seven and two class classification. The duo (ResNet50 and Levenberg Marquardt) produced a classification accuracy of 82.92%. Among all the classes of diagnosis, class 7 has the best F-value followed by class 1, whereas class 4 has the lowest F- value followed by class 5 and class 2. Lowest F-value indicates maximum misclassification. For two-class classification, duo (ResNet50 and Levenberg Marquardt) produced classification accuracy is 94.77%. The F-value of both the classes is above 92% for all the combination of CNN and shallow neural network.

The results do conclude that the deep neural networks can easily classify the cases of cervical cancer with notable accuracy, without feature extraction.