Announcement of Official Publication of Paper on AI for Breast Carcinoma Detection in Histopathological Images during Intraoperative Rapid Diagnosis

We are engaged in research and development aimed at solving medically and socially important issues, including Japan’s declining birthrate and aging population, and our research also targets disease areas affecting women and children.
As part of these efforts, through joint research with Professor Takashi Suzuki and Assistant Professor Mio Yamaguchi-Tanaka of the Department of Pathology and Histotechnology, Tohoku University Graduate School of Medicine (National University Corporation Tohoku University), and NEC Solution Innovators, Ltd. (NES), we have been developing artificial intelligence (AI) that, using frozen pathological sections1) of breast cancer, supports intraoperative rapid diagnosis2).
The paper summarizing the results of this research, “Artificial Intelligence for Breast Carcinoma Detection in Histopathological Images Based on Single Shot MultiBox Detector in Intraoperative Rapid Diagnosis,” was announced in our news release of December 5, 2025, following its advance online publication on J-STAGE on December 4, 2025.
We are pleased to announce that the final journal version of the paper has now been officially published in The Tohoku Journal of Experimental Medicine, Volume 269, Issue 4 (August 2026 issue).
Tohoku J. Exp. Med., 2026, 269(4), 507–516.
DOI: https://doi.org/10.1620/tjem.2025.J120
News release of December 5, 2025 (in Japanese): https://www.renascience.co.jp/breast-cancer-ai/
Please note that, although bibliographic details and notation have been refined in the final published version, there are no changes to the content of the study or its main results compared with the advance online version.

Background of the Joint Research
Breast cancer is the most common cancer among Japanese women, and it is said that one in eleven Japanese women will develop breast cancer in her lifetime. When breast cancer is suspected based on a lump or imaging findings, the final diagnosis is made by pathological diagnosis. To provide prompt and appropriate treatment for breast cancer, an accurate diagnosis by pathologists based on histological classification is essential.
However, the number of pathologists is limited, and expectations are growing for AI technologies that support diagnosis.
Together with Tohoku University and others, we are developing an AI that detects breast cancer lesions in pathological images. On October 7, 2022, we announced that the results of research on a breast cancer detection model for histopathological micrographs4), using the Single Shot MultiBox Detector (SSD)3), one of the AI-based object detection methods, had been published in the scientific journal The Journal of Pathology Informatics.
In that study, when breast cancer was classified into 3 classes (“benign, non-invasive carcinoma5), invasive carcinoma6)”) or 2 classes (“benign, malignant”), diagnosis was shown to be possible with accuracies of 88.3% and 90.5%, respectively.
In the present study, we applied this technology to intraoperative rapid pathological diagnosis of breast cancer.
Intraoperative rapid diagnosis is an important pathological diagnosis in breast cancer surgery for determining matters such as the extent of resection. On the other hand, it must be performed with limited time and personnel, and pathologists with a high level of expertise in breast cancer pathology are required. Furthermore, in addition to the need to make a diagnosis within 10–20 minutes of receiving the specimen, frozen sections prepared during surgery tend to be of lower quality than ordinary pathological specimens.
In light of these challenges, in this study we worked on developing AI that supports intraoperative rapid diagnosis.

Results
After training the SSD model using microscopic images of 943 intraoperative frozen sections of breast cancer as training data, we evaluated diagnostic accuracy for breast cancer using 65 intraoperative frozen section images.
As a result, the following performance was obtained:
• Classification of breast cancer: Distinguishing benign from malignant, and further classifying malignant cases into non-invasive carcinoma and invasive carcinoma
• Diagnostic accuracy:
– Benign / malignant classification: accuracy of 92.3%
– Classification into non-invasive carcinoma / invasive carcinoma in addition to benign / malignant: accuracy of 89.2% (with the detection confidence threshold optimized for each class)
• Average detection time: 0.875 seconds per image
These results indicate the potential for this AI to become a technology that supports intraoperative rapid diagnosis of breast cancer.

Outlook
The diagnostic technology using this AI may become a tool that supports intraoperative rapid diagnosis of breast cancer, thereby reducing the workload of pathologists and supporting surgical decision-making based on intraoperative diagnosis.
On the other hand, toward future clinical application, further expansion of the training data and optimization of detection conditions remain challenges.
Positive breast cancer cases obtained in intraoperative rapid diagnosis are limited, and cases of invasive carcinoma are particularly few; the number of invasive carcinoma images collected in this study was 45 across the entire dataset. Expanding the training data in the future may improve the detection performance for invasive carcinoma.
In addition, the dataset in this study includes both ductal carcinoma and lobular carcinoma. The paper also discusses the possibility that detection performance could be further improved by securing a sufficient number of images and training the model to distinguish between them.
One way to prevent AI from overlooking cancer cells as much as possible is to lower the detection threshold to increase recall. On the other hand, if the information presented by the AI becomes excessive, it may complicate pathologists’ decisions and could instead increase their burden.
Therefore, setting conditions that appropriately adjust the amount of information presented to pathologists while minimizing oversight of cancer cells is an important issue for future clinical application.
We hope that this model will develop into a technology that supports pathologists’ diagnosis and contributes both to reducing their workload and to lowering the risk of overlooking cancer cells in intraoperative diagnosis.

1) Frozen pathological sections
Specimens prepared by rapidly freezing tissue taken from a tumor such as breast cancer during surgery, slicing it thinly, and observing it under a microscope. They are used to quickly confirm during surgery whether the resection range of the tumor is sufficient and whether there is lymph node metastasis, and to decide surgical policy, such as additional resection, as needed.

2) Intraoperative rapid diagnosis
A method of preparing frozen sections from tissue removed during surgery and examining them under a microscope in a short time to diagnose the presence or absence of cancer.

3) Single Shot MultiBox Detector (SSD)
One of the AI technologies that detects the position and type of objects in an image. In this study, it is used to detect and classify breast cancer lesions from microscopic images of intraoperative frozen sections.

4) Histopathological micrographs
Images obtained by thinly slicing part of breast cancer tissue or lymph nodes removed by surgery, staining it, and photographing it under a microscope. They allow confirmation of the presence and morphology of cancer cells, their state of growth, and their spread into surrounding tissue, and provide important information for deciding treatment policy.

5) Non-invasive carcinoma
Cancer in which the cancer cells have not invaded the surrounding stroma beyond the basement membrane. Although the risk of metastasis to surrounding tissue is low, it is important in intraoperative diagnosis to confirm the extent and boundaries of the lesion.

6) Invasive carcinoma
Cancer in which the cancer cells have invaded the surrounding stroma beyond the basement membrane. There is a possibility of invasion into blood vessels and lymphatic vessels and of metastasis, which is important in determining the surgical method and extent of resection.