Impact of pre‐analytic variables on deep learning accuracy in histopathology

Abstract

Aims

Machine Learning (ML) binary classification in diagnostic histopathology is an area of intense investigation. Several assumptions, including training image quality/format and the number of training images required, appear similar in many studies irrespective of the paucity of supporting evidence.

Methods

We empirically compared training image file type, training set size, and two common Convolutional Neural Networks (CNN) using transfer learning (ResNet50, SqueezeNet). Thirty H&E slides with carcinoma or normal tissue from three tissue types (breast, colon, prostate) were photographed generating 3,000 partially overlapping images (1,000 per tissue type). These lossless PNGs were converted to lossy JPGs. Tissue type‐specific binary classification ML models were developed using all PNG or JPG images, and repeated with a subset of 500, 200, 100, 50, 30, and 10 images. Eleven models were generated for each tissue type, at each quantity of training images, for each file type, and for each CNN, resulting in 924 models. Internal accuracies and generalization accuracies were compared.

Results

There was no meaningful significant difference between accuracies in PNG vs JPG models. Models trained with more images did not invariably perform better. ResNet50 typically outperformed SqueezeNet. Models were generalizable within a tissue type but not across tissue types.

Conclusions

Lossy JPG images were not inferior to lossless PNG images in our models. Large numbers of unique H&E slides were not required for training optimal ML models. This reinforces the need for an evidence‐based approach to best practices for histopathologic ML.

This article is protected by copyright. All rights reserved.

https://ift.tt/2TZCFz7