Application of the sliding window method and Mask-RCNN method to nuclear recognition in oral cytology

Diagnostic Pathology

Table 1 Results of the sliding window method (SWM)

No	Activation function	Epoch	No of layers	Dropout	Value of the training loss function	Percentage of correct answers for the training group	Value of the loss function of the test group	Percentage of correct answers in the test group
1	ReLU	5	18	0.25	0.3356	0.8668	0.3266	0.8654
2	ReLU	20	18	0.25	0.0891	0.9562	0.2104	0.9314
3	ReLU	50	18	0.25	0.0526	0.9851	0.6082	0.9111
4	ReLU	5	24	0.25	0.351	0.8524	0.3188	0.8705
5	ReLU	20	24	0.25	0.0798	0.9251	0.1958	0.9314
6	ReLU	50	24	0.25	0.0628	0.9982	0.4614	0.8908
7	ReLU	5	18	0.5	0.3752	0.8226	0.3225	0.8705
8	ReLU	20	18	0.5	0.2151	0.9198	0.2514	0.9213
9	Sigmoid	20	18	0.25	0.7456	0.4659	0.7285	0.4785

The results obtained using the SWM and convolutional neural network (CNN) methods are shown. To improve the accuracy of the loss function values and correctness of the answer rate, several dropouts were attempted to increase the number of epochs and avoid overlearning. We also deepened the network by increasing the number of convolution, pooling, and activation functions
The CNN with the highest number of correct answers in the training data was No.6 (99.8%), whereas the percentage of correct answers in the test data was 89.1%. This is thought to be due to overlearning. The CNN with the highest accuracy and without overlearning was No. 2 (93.1% correct answers on the test data)

ISSN: 1746-1596