Development of a Deep learning-based pipeline to classify Small Round Cells Sarcomas histotypes

Ewing sarcoma (ES), Ewing-like sarcomas (ELS) and undifferentiated synovial sarcoma (SS) represent the main entities belonging to the family of the Small Round Cell Sarcomas (SRCS), a group of rare, heterogenous and highly aggressive mesenchymal tumors. SRCS are classified according to a specific single gene rearrangement. However, despite specific histological features are strongly correlated with the underlying molecular alteration, morphological overlapping may occur, and combined with their rarity, make the diagnosis challenging especially for non-expert pathologists. Within this context, the spreading of digital pathology and the recent developments of deep learning technologies for image processing, offer new opportunities for analysis, interpretation, and classification of histopathological slides. In this study, a deep learning-based framework called DeeRasNET, is specifically developed to classify hematoxylin and eosin-stained slides of ES, SS, BCOR and CIC rearranged sarcomas. Accuracy was the main metrics parameter used to evaluate the model performance. Initially, due to the small size of the datasets implemented for the model training, the classification accuracy for each class of sarcoma resulted low (mean accuracy of 0.6). To increase the performance of the model, we developed a pre-processing semi-automated pipeline comprising an open-source graphical interface unit (called TilerPath) with which we managed the tissue whole slide images, selecting interesting tissue areas and performing a quality control of the images used for classifier implementation. By TilerPath uninformative and misleading images were excluded from the model. After pre-preprocessing by Tilerpath, a total of 18193 tiles, selected from 124 digital slides covering all the four histotypes investigated, was used to train and test DeeRasNET. Finally, the scalability of the system was demonstrated on a validation dataset comprising 2706 tiles randomly selected from cases not included into the training and test set. After quality improvement, the final model showed a strong increase of classification performance, with accuracies ranging from 0.98 to 0.99 among all the sarcoma types. Both the TylerPath and the DeeRASnet source code were released as open-source software.