Gastric cancer remains one of the leading causes of cancer and death worldwide. According to GLOBOCAN 2020 global cancer statistics, gastric cancer contributed to 5.6% of all new cancer cases and 7.7% of all cancer-related deaths (Sung et al., 2021), which made it the fifth most common cancer and ranked the third leading cause of death worldwide (Ajani et al., 2022). East Asia has the world's highest incidence of gastric cancer, and it is the third most common new cancer case in China (Xia et al., 2022).

Gastric cancer is often in advanced stages when first detected. It has a heterogeneous nature that can present with different histological subtypes, categorised into intestinal, diffuse, and mixed types based on Lauren's criteria (Berlth et al., 2014). Each cancer type has distinct morphological and molecular features, e.g. adenoid differentiation in the intestinal type, and irregular and diffused structure in the diffuse type (Berlth et al., 2014). Mixed-type gastric cancer, in particular, is characterised by the co-existence of glandular and undifferentiated components and is associated with a more unfavourable prognosis compared to the other two subtypes. Pathological assessment of tissue specimens continues to serve as the gold standard for diagnosis, yet it heavily relies on the expertise of pathologists and their ability to accurately identify the glandular and undifferentiated components in the tissue sections.

As the precise diagnosis of mixed-type gastric cancer can be challenging due to its complex histological features and overlapping characteristics with other gastric cancer subtypes, Artificial Intelligence may have transformative potential to overcome this clinical challenge. Here, we investigated the feasibility of machine learning techniques, which have been fast developed in clinical oncology for diagnosis, predicting prognosis and informing clinical decisions (Swanson et al., 2023). The convolutional neural network is the most commonly used approach that has been implemented in cancer diagnosis using endoscopic images (Hirasawa et al., 2018). U-Net is one such model for image segmentation, which uses a U-shaped topology with an encoder and a decoder to extract and analyse the features of the image through convolution and pooling to generate segmentation results ($author1$ et al., 7] </id><AuthGrp><Author><au>Ronneberger</au><dl> </dl><in>O</in></Author><dl>, </dl><Author><au>Fischer</au><dl> </dl><in>P</in></Author><dl>, </dl><Author><au>Brox</au><dl> </dl><in>T.</in></Author></AuthGrp> <subtitle>U-Net</subtitle>: Convolutional Networks for Biomedical Image Segmentation; proceedings of the Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, <misc>Cham, F 2015//, 2015 [C]. Springer International Publishing</misc>.). To date, U-Net has been widely used in developing automated medical image segmentation, especially in cancer images, with high accuracy and robustness (Yin et al., 2022, Wang et al., 2021); however, its application in diagnosing mix-type gastric cancer from pathological images has not been reported. On the other hand, QuPath is an open-source pathology software for image analysis, including visualisation, segmentation, classification, and quantitation (Bankhead et al., 2017), which has a strong potential for clinical use in the future. Therefore, we trained both QuPath and U-Net to automate the segmentation of undifferentiated and differentiated components in pathological images of mixed-type gastric cancer.

In recent years, mixed-type gastric cancer has been found to be associated with higher risks of lymph node metastasis at both early and advanced stages compared with the other two types (Horiuchi et al., 2020, Lu et al., 2021). The metastatic risk may be closely related to the differentiation status of the cancer cells rather than the staging, where undifferentiated components have a higher likelihood of migrating to the lymph nodes and thereafter, remote organs (Takeuchi et al., 2018). The prediction of lymph node metastasis is crucial for deciding the extent of lymphadenectomy by surgeons. Insufficient lymphadenectomy may result in cancer recurrence and metastasis, whereas over-lymphadenectomy can lead to complications, such as oedema and lymphatic fistula, affecting post-operational recovery (Li et al., 2018). Therefore, in this study, we correlated the ratio of differentiated/undifferentiated components to lymphatic metastasis identified during surgery to determine whether this ratio is a good predictor for the risk of metastasis of mixed-type gastric cancer during treatment planning.