Artificial Intelligence for Breast Carcinoma Detection in Histopathological Images Based on Single Shot Multibox Detector in Intraoperative Rapid Diagnosis

Intraoperative rapid diagnosis with pathological classification by pathologists plays a pivotal role in making appropriate surgical decisions.However, morphological evaluation of intraoperative frozen sections is apparently more difficult than that of formalin fixed paraffin embedded (FFPE) tissue because of their unavoidable tissueprocessing artifacts, which compels pathologists, even well-trained, more challenging to make an accurate diagnosis.Therefore, assisting the pathological diagnosis with artificial intelligence is anticipated.Here, we built the data set consisting of histopathological micrographs of intraoperative frozen sections and demonstrated an automatic breast carcinoma detection model based on a Single Shot Multibox Detector, an object detection method known for real-time, high-speed performance.The final 3-class classification model was trained on 943 intraoperative frozen section images and evaluated using 65 images, identifying benign lesions, non-invasive carcinoma, and invasive carcinoma.The 2-class (benign or malignant) and 3-class classification tasks achieved diagnostic accuracies for whole image of 92.3% and 89.2%, respectively.When we evaluated the performance of object detection in the 2-class task, the precision, recall and F1 scores for malignant detection are 70.3%, 74.4%, and 72.3%, respectively.Additionally, the model's object detection performance was assessed using FFPE images, yielding a precision of 55.1%, a sensitivity of 76.9%, and an F1 score of 64.2%, suggesting that it is particularly well-suited for intraoperative frozen sections.In conclusion, we presented the automatic breast carcinoma detection method, utilizing histopathological micrographs of intraoperative frozen sections.The present model may assist pathologists by reducing their workload and minimizing the risk of overlooking carcinoma cells in intraoperative diagnosis.