Miniature probability maps using resource limited embedded device for classification of histopathological images

Abstract

Prediction of malignancy in histopathology images using CNN is mostly performed using cloud services suffers from network latency. We propose a novel, efficient method to classify whole slide histopathology images using modular and portable embedded devices to detect the presence of cell abnormality. The proposed method generates probability maps which indicates predictions so that a bird’s-eye view of tissue malignancy can be obtained. The miniature map(mini-map) of histopathology image is the overview of binary class probabilities at the patient level. The computational overhead of device is reduced as well as prediction will be faster while using custom-trained model. The round trip time is also reduced as the computing occurs near the end-device itself. The obtained predictions in mini-map can be viewed in any portable device consuming minimum processing time as the size of the map is only few kilo-bytes. This method is found to be suitable to assist medical practitioners in patient diagnosis.