RESUMO
Around 1% of the population of the UK and North America have a diagnosis of coeliac disease (CD), due to a damaging immune response to the small intestine. Assessing whether a patient has CD relies primarily on the examination of a duodenal biopsy, an unavoidably subjective process with poor inter-observer concordance. Wei et al. [11] developed a neural network-based method for diagnosing CD using a dataset of duodenal biopsy whole slide images (WSIs). As all training and validation data came from one source, there was no guarantee that their results would generalize to WSIs obtained from different scanners and laboratories. In this study, the effects of applying stain normalization and jittering to the training data were compared. We trained a deep neural network on 331 WSIs obtained with a Ventana scanner (WSIs; CD: n=190; normal: n=141) to classify presence of CD. In order to test the effects of stain processing when validating on WSIs scanned on varying scanners and from varying laboratories, the neural network was validated on 4 datasets: WSIs of slides scanned on a Ventana scanner (WSIs; CD: n=48; normal: n=35), WSIs of the same slides rescanned on a Hamamatsu scanner (WSIs; CD: n=48; normal: n=35), WSIs of the same slides rescanned on an Aperio scanner (WSIs; CD: n=48; normal: n=35), and WSIs of different slides scanned on an Aperio scanner (WSIs; CD: n=38; normal: n=37). Without stain processing, the F1 scores of the neural network were 0.947, 0.619, 0.746, and 0.727 when validating on the Ventana validation WSIs, Hamamatsu and Aperio rescans of the Ventana validation WSIs, and Aperio WSIs from a different source respectively. With stain normalization, the performance of the neural network improved significantly with respective F1 scores 0.982, 0.943, 0.903, and 0.847. Stain jittering resulted in a better performance than stain normalization when validating on data from the same source F1 score 1.000, but resulted in poorer performance than stain normalization when validating on WSIs from different scanners (F1 scores 0.939, 0.814, and 0.747). This study shows the importance of stain processing, in particular stain normalization, when training machine learning models on duodenal biopsy WSIs to ensure generalizability between different scanners and laboratories.
RESUMO
We present a multiple-instance-learning-based scheme for detecting coeliac disease, an autoimmune disorder affecting the intestine, in histological whole-slide images (WSIs) of duodenal biopsies. We train our model to detect 2 distinct classes, normal tissue and coeliac disease, on the patch-level, and in turn leverage slide-level classifications. Using 5-fold cross-validation in a training set of 1841 (1163 normal; 680 coeliac disease) WSIs, our model classifies slides as normal with accuracy (96.7±0.6)%, precision (98.0±1.7)%, and recall (96.8±2.5)%, and as coeliac disease with accuracy (96.7±0.5)%, precision (94.9±3.7)%, and recall (96.5±2.9)% where the error bars are the cross-validation standard deviation. We apply our model to 2 test sets: one containing 191 WSIs (126 normal; 65 coeliac) from the same sources as the training data, and another from a completely independent source, containing 34 WSIs (17 normal; 17 coeliac), obtained with a scanner model not represented in the training data. Using the same-source test data, our model classifies slides as normal with accuracy 96.5%, precision 98.4% and recall 96.1%, and positive for coeliac disease with accuracy 96.5%, precision 93.5%, and recall 97.3%. Using the different-source test data the model classifies slides as normal with accuracy 94.1% (32/34), precision 89.5%, and recall 100%, and as positive for coeliac disease with accuracy 94.1%, precision 100%, and recall 88.2%. We discuss generalising our approach to screen for a range of pathologies.
