Automatic discrimination between neuroendocrine carcinomas and grade 3 neuroendocrine tumors by deep learning of H&E images

Neuroendocrine neoplasms (NENs) arise from diffuse neuroendocrine cells and are categorized as either well-differentiated and less proliferative Neuroendocrine Tumors (NETs), divided into low (G1), middle (G2), and high grades (G3), or poorly differentiated, and more proliferative Neuroendocrine Carcinomas (NECs). Low-grade NENs typically necessitate surgical intervention, whereas high-grade ones often require chemotherapy. However, low-grade NENs may exhibit aggressive behavior. Therefore, it is crucial to precisely refine the diagnosis of NENs. This refinement is achievable when differentiation/non-differentiation is evident or when the Ki-67 or mitosis index is low. The challenge arises in cases of morphologically undifferentiated instances with a high Ki-67 percentage and/or high mitotic index. To address this challenge, we developed a Deep Learning (DL) system named NEToC, designed to differentiate between NETs and NECs using exclusively morphological information from immunohistochemistry images, without relying on Ki-67 or mitosis assessments. NEToC was developed using 95 NEN cases from the period 2015 to 2018 at Parc Tauli Hospital in Spain, comprising 588 images. Implemented as a Graphical User Interface (GUI) system, NEToC is intended for deployment in pathological departments of hospitals to perform federated supervision. We tested the performance of NEToC with 119 images that were not used during the Artificial Neural Network (ANN) training phase, and evaluated its robustness across various resolutions: 64 × 64, 128 × 128, 256 × 256, and 512 × 512 pixels. The achieved accuracies for these resolutions were 74 %, 98 %, 98 %, and 100 %, respectively, for an underrepresented NET G3 experiment, and 66 %, 89 %, 95 % and 94 % for a represented NET G3 experiment. Based on several measured performance metrics, the optimal resolution appears to be between 128 × 128 and 256 × 256 pixels, considering computational resources and accuracy requirements. However, we found that the 256 × 256-pixel resolution is more robust to classify underrepresented classes in the learning phase. These results imply that the information to discriminate between NECs and Grade 3 NETs needs to be resolved in regions with a pixel resolution of no more than 4 μm/pixel. Most of the misclassifications were false negatives, where NET G1-type images were erroneously classified as NEC-type. Our results demonstrate that a DL-based diagnostic algorithm provides a more accurate diagnosis in NEN cases where physicians face challenges. NEToC has been initially trained with and used to classify gastrointestinal NENs. Since the NEN morphology does not change among the different organs, the use of NEToC can be extrapolated to NENs from different organs. NEToC facilitates federated supervision, allowing pathologists to collect interchangeable files based on NEToC classification predictions. NEToC is an easy-to-use, adaptable software that integrates multiple ANNs to improve standardization and accuracy NEN diagnosis, opening up possibilities for combining DL and histological diagnosis in federated supervision systems. A future goal is to classify not only NETs, but also the three-tier system (NET G1, NET G2, and NET G3) based solely on tissue differentiation information.