DIAGNOSIS AND CLASSIFICATION ASSISTANCE FROM LYMPHOMA MICROSCOPIC IMAGES USING DEEP LEARNING
Pierre Brousset, Charlotte Syrykh, Arnaud Abreu, Nadia Amara, Camille Laurent · Hematological Oncology · 2019
The microscopic diagnosis of lymphoma remains challenging. Recent data from our group within the French (nationwide) Lymphopath network shows that 20% of diagnoses are inaccurate, with direct impact on patient care1. Another difficulty is the subclassification of lymphoma subtypes to predict therapeutic response and clinical behaviour. For these, molecular techniques, not affordable for all pathology departments, have become critical. Currently, automated solutions that could help with diagnostic decisions or histological grading are lacking, and it is unreasonable to expect pathologists to be experts on rare tumours if they only see a few cases per year. Digital microscopy offers unique features which are not available in conventional optical microscopy2 and allows automated image analysis and quantification by using computer vision and, in particular, deep learning (DL) approaches.3 Using DL on digital slides we tried to address two questions: 1) Can we train deep neural network (CNN) to distinguish follicular lymphoma (FL) from either follicular hyperplasia or other small B-cell lymphoma subtypes on hematoxylin-eosin stained lymph node digital slides? 2) Can CNN separate germinal center from non germinal center diffuse large B-cell lymphoma (DLBCL) the same way? Patch extraction or patch coordinated extraction was carried with different architectures (Residual Net, GoogleNet, VG-GNet) of whole slides images. A home designed neural network classification was also performed on 125x125pixel non-overlapping tiles of a whole slide image extracted at very low resolution (pyramid level 5 ó 7,68μm/pixel). The latter classifier was a fairly simple CNN architecture. It stacked three blocks of Convolution-MaxPooling with respectively 32, 64 and 128 filters, followed by two 1024-unit fully connected layers. Output of the network is a 2-unit layer with softmax activation to produce class probabilities. As to the diagnosis of FL, the accuracy of our method is superior to that of non expert pathologists, with an average area under the curve (AUC) of 0.95. Furthermore, we trained CNNs to predict GC and non GC phenotype of DLBCL previously determined with Hans' algorithm with an average AUC of 0.882. Our models were validated on independent series of biopsies (training, testing and validation sets) and provide quite similar accuracy irrespective of the architecture we used. These findings strongly suggest that DL models, even on very simple CNNs, can assist pathologists in the diagnosis and subclassification of lymphoma as recently shown is subsets of lung cancers.3 References 1) Laurent C et al. J Clin Oncol. 2017;35:2008-17. 2) Krenacs T et al. Digital microscopy - 2010. Microscopy: Science, Technology, Applications and Education. A. Méndez-Vilas and J. Díaz (Eds.). 3) Coudray N et al. Nat Med. 2018;24(10:1559-67). Keywords: diffuse large B-cell lymphoma (DLBCL); follicular lymphoma (FL).