Performance of a machine learning model in recognition of normal tissue from histological sections

Abstract Introduction Machine learning and artificial intelligence (AI) models have been applied in histopathology to solve specific problems like detection of metastasis in lymph nodes and immunohistochemical scoring. We have aimed to develop a machine learning model which can be trained in histopathology from the basics, i.e. identification of normal tissue. We have tried to replicate the process through which a human pathologist learns recognition of normal tissue from histological sections, and evaluate the performance of a machine learning model at this task. Materials and methods A total of 658 histologic images were anonymised, microphotographed at 10x magnification, under the same condition of illumination, with a Magnus DC5 integrated microphotography system. The images were split into two subsets, training (386) and validation (272 images). The images belonged to seven classes of tissue: brain, intestine, kidney, liver, lungs, muscle and skin. Archived material of the hospital were used for the study. A machine learning model using convolutional neural network (CNN) was developed on the Keras platform, using the convolution layers of a pretrained VGG16 model. The model was trained with the training set of images over 10 epochs. After training, performance of the model was assessed on the validation set. Results The model achieved 88.24% accuracy in classifying the images of the validation set. The most frequent errors were met in recognising images of kidney (14 errors, 33.33%). The commonest error was wrongly classifying kidney tissue as liver (07 errors). Analysis of the deeper layers of the neural network revealed specific patterns in images which were wrongly classified. Conclusion The results of the present study indicates that a convolutional neural network might be trained in histology similar to a trainee pathologist. The study represents the first step towards developing a machine learning model as a generalised histopathological image classifier.