Accurate preoperative classification of gastric cancer (GC) is essential for determining appropriate treatment strategies. The Lauren classification system categorises GC into intestinal and diffuse types, influencing both prognosis and therapeutic approaches. While conventional diagnostic methods, such as endoscopic biopsy, provide histological insights, they have limitations, including invasiveness and potential sampling errors. Noninvasive alternatives are being explored, with radiomics—extracting quantitative features from medical images—emerging as a promising approach. A recent study published in Insights into Imaging evaluates an automated deep learning method that combines nnU-Net with radiomics to classify GC preoperatively, potentially improving accuracy and efficiency.
Automated Segmentation Using nnU-Net
The study included 433 GC patients from three medical centres, with a subset used to train an nnU-Net model for automatic tumour segmentation. The nnU-Net framework, based on the U-Net architecture, was specifically adapted to identify GC lesions on computed tomography (CT) images. The model demonstrated high segmentation accuracy, achieving a Dice similarity coefficient (DSC) of 0.79 in the test set. This result highlights the capability of nnU-Net to provide efficient and precise tumour delineation. Automated segmentation minimises manual workload and subjectivity, facilitating more consistent and efficient radiomic feature extraction. By eliminating manual segmentation inconsistencies, nnU-Net enhances reproducibility and reliability in clinical applications. The reliance on automated segmentation reduces the variability often introduced by manual methods, ensuring that extracted tumour characteristics remain consistent across different patients and imaging datasets. This consistency is particularly important in multicentre studies, where variations in imaging protocols can affect diagnostic accuracy.
Radiomics-Based Classification Model
Radiomics techniques were applied to the segmented tumour regions, extracting 1,874 quantitative features, including first-order statistical properties, shape attributes and texture-based characteristics. These features capture important tumour characteristics that may not be visible to the human eye. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression method to identify the most relevant predictors for classification.
Recommended Read: Advancing Gastric Cancer Imaging with [18F]ALF-NOTA-FAPI-04 PET/CT
A logistic regression model trained on these optimised features demonstrated strong predictive capabilities, achieving area under the curve (AUC) values ranging from 0.84 to 0.86 across various validation datasets. The model effectively distinguished between intestinal and diffuse GC subtypes, reinforcing the potential of radiomics in noninvasive cancer classification. By leveraging advanced feature selection techniques, the model ensures robust performance, reducing redundancy in the dataset and focusing on the most informative characteristics for classification. Radiomics offers a major advantage over traditional histopathological approaches by providing a holistic assessment of tumour heterogeneity. Unlike biopsies, which sample only a small portion of the tumour, radiomics analyses the entire lesion, offering a more comprehensive representation of tumour biology.
Integration of Clinical and Radiomic Data
In addition to radiomic features, clinical variables such as patient age, CA125 levels and tumour diameter were incorporated into a combined prediction model. However, comparative analyses revealed no statistically significant performance improvement over the standalone radiomic model. This suggests that radiomics alone provides a reliable classification framework without additional clinical variables. The findings indicate that radiomic data capture essential tumour characteristics necessary for classification, potentially simplifying clinical decision-making.
The study highlights the robustness of radiomics for preoperative GC classification and supports its integration into clinical workflows. By automating the classification process, this approach could assist clinicians in decision-making while reducing the dependence on invasive diagnostic procedures. Moreover, the study underscores the growing role of artificial intelligence in oncological diagnostics, demonstrating that machine learning models can rival traditional clinical assessment methods in accuracy. The ability to accurately classify GC subtypes using noninvasive imaging techniques opens the door to personalised medicine, where treatment plans can be tailored to individual patient profiles without requiring tissue samples.
The combination of nnU-Net and radiomics represents a promising approach for the preoperative classification of GC based on the Lauren subtypes. The automated segmentation provided by nnU-Net enhances efficiency and reproducibility, while radiomic features enable accurate and noninvasive classification. The study’s findings emphasise the potential of deep learning-driven radiomics to support precision oncology, offering a reliable alternative to traditional diagnostic methods. By reducing reliance on invasive procedures, this approach allows for more personalised treatment planning, ultimately improving patient outcomes. The integration of these advanced computational techniques into clinical practice could revolutionise GC management, paving the way for more efficient, cost-effective and accessible cancer diagnostics. As healthcare systems increasingly adopt AI-driven solutions, the role of radiomics in preoperative cancer classification is likely to expand, further enhancing its value in clinical settings.
Source: Insights into Imaging
Image Credit: iStock
