Artificial Intelligence in Gastric Cancer Imaging With Emphasis on Diagnostic Imaging and Body Morphometry

T-staging

For T-staging of gastric cancer, radiologists assessed the presence of focal or diffuse wall thickening or mass-like lesions with densities different from those of the normal stomach wall on contrast-enhanced CT images. The detailed criteria for CT staging of the T stage are summarized in Fig. 1 [30]. Even experienced radiologists often encounter challenges in accurately distinguishing between EGC (T1 cancer) and T2 stage gastric cancer on CT scans. The accuracy for this differentiation ranges from 51.6% to 91.5%. Furthermore, discriminating between T1a and T1b stage cancers is more challenging, with accuracies ranging from 62.5% to 69.2% [31]. Zeng et al. [9] reported that AI classifier models based on ResNet101 demonstrated high accuracy in distinguishing between EGC (T1 cancer) and T2 cancer, with accuracies ranging from 91.4% to 94.6%. Furthermore, their model demonstrated the ability to discriminate between T1a and T1b cancers with accuracies ranging from 62.3% to 88.6% [9]. Despite these promising results, these AI models have not yet been implemented in clinical practice owing to their insufficient diagnostic accuracy compared to endoscopic ultrasonography, which remains the preferred method for evaluating the depth of invasion.

Fig. 1
Criteria for computed tomography staging of gastric cancer with sample cases.
LN = lymph node.

• Click for larger image
• Download as PowerPoint slide

Another challenge in T-staging using CT is the evaluation of serosal invasion in gastric cancer. The accuracy of CT in identifying serosal invasion has been reported to be 76.6% (sensitivity,96%; specificity,72.4%) [10]. As T4a stage cancer with serosal invasion is considered to increase the risk of peritoneal metastases, identifying serosal invasion is important for treatment planning. Recently, Sun et al. [10] reported that a CT-based deep learning radiomics model could accurately evaluate serosal invasion in AGC, with area under curves (AUCs) ranging from 0.76 0.90. However, these AI radiomics models require the manual segmentation of tumors, which is time-consuming. In addition, these models have not been validated in many external validation cohorts; thus, they have not yet been clinically applied.

N-staging

Lymph node (LN) metastasis is a significant prognostic factor for gastric cancer. The accurate assessment of LN metastasis before surgery is crucial for making treatment decisions, such as determining the suitability of endoscopic mucosal resection, neoadjuvant chemotherapy, or gastrectomy with or without lymphadenectomy. Currently, the commonly used criteria for N-staging on contrast-enhanced CT are as follows: 1) LNs with a short-axis diameter ≥8 mm; 2) a cluster of 3 or more peritumoral LNs regardless of size; 3) LNs showing strong enhancement >100 HU; or 4) LNs with central necrosis and perinodal infiltration regardless of size [32]. However, N-staging on contrast-enhanced CT scans is unsatisfactory owing to its low diagnostic accuracy, leading to the development of several AI techniques with or without radiomics for evaluating LN metastasis [19, 20, 21, 22]. These studies have utilized both radiomics analysis of handcrafted primary gastric cancer lesions or metastatic LNs as well as deep learning models.

Most studies have focused on distinguishing between the N+ (LN-positive) and N- (LN-negative) statuses [19, 21, 22]. However, Dong et al. [20] aimed to develop a deep learning radiomic nomogram to predict the number of LN metastases. The model in the study by Dong et al. [20] was a well-validated international multicenter collaboration, which was tested in three external validation cohorts and one international validation cohort, ensuring the reliability and reproducibility of their results across different centers. They demonstrated that the deep learning radiomic nomogram achieved good discrimination in predicting the number of metastatic LNs across all cohorts, with C-indices ranging from 0.797 to 0.821 [20]. However, these advanced models require manual tumor region segmentation by an experienced radiologist, which limits their practical utility in daily clinical practice.

M-staging

The evaluation of peritoneal metastasis is the most crucial aspect of M staging in gastric cancer. Typically, CT images show specific characteristics associated with peritoneal metastasis, including peritoneal fat stranding, omental cakes, ascites, parietal peritoneal thickening, and nodules/masses. However, when these features are subtle, radiologists may overlook them, leading to reduced sensitivity and diagnostic accuracy of peritoneal metastases. In addition, occult peritoneal metastasis (OPM) refers to lesions in which the initial CT scan indicates negative peritoneal metastasis but subsequent laparoscopy or surgery reveals its presence. Owing to the inherent nature of OPM, CT scans have limited sensitivity and diagnostic accuracy in patients with OPM [23]. Consequently, extensive research efforts have focused on developing radiomics or AI models to enhance the detection rates of peritoneal metastasis and OPM on CT scans [23, 24, 25, 26].

Previous studies have predominantly employed radiomics analysis based on manually crafted features from primary gastric cancer or potential peritoneal lesions, yielding AUC values ranging from 0.724 to 0.836 [24, 25, 26]. However, one study utilized a deep learning model that exhibited improved diagnostic accuracy with an AUC of 0.900, surpassing that of the conventional clinical model (AUC of 0.670) [23]. Nevertheless, a common challenge shared by all M-staging studies, as well as T-staging and N-staging, is the need for manual segmentation of tumors, which significantly hampers their practical application in routine clinical practice.

Algorithms for sarcopenia evaluation

According to the revised European Working Group on Sarcopenia in Older People (EWGSOP2), the muscle area measured at the third lumbar vertebral level on CT was used as a representative value because it reflects overall body muscle mass [33]. Deep learning techniques enable automatic segmentation of the muscle and fat areas. The authors developed a fully automated AI technique for body morphometry, consisting of 2 AI algorithms: 1) automatic selection of the L3 vertebral level and 2) automated segmentation of the muscle and fat areas [3, 13]. The overall framework of the fully automated body morphometry is illustrated in Fig. 2. This deep learning model offers a fully automated selection of the axial CT slice at the L3 vertebral level and accurate segmentation of the abdominal muscle area, irrespective of the anatomical variations. The average cross-sectional area errors of the deep learning model were 2.22% in the normal anatomy group and 2.37%–4.06% in the anatomic variation group [13]. Recently, this AI body mophometry solution (AID-U™; iIAD Inc., Seoul, Korea) obtained regulatory approval, enabling its utilization in clinical practice.

Fig. 2
Framework of the artificial intelligence techniques for fully automated body morphometry.
CT = computed tomography; L3 = the third lumbar spine vertebra.

• Click for larger image
• Download as PowerPoint slide

Algorithms for myosteatosis evaluation

Recently, there has been a growing interest in evaluating myosteatosis, which refers to excessive fat infiltration within skeletal muscles, as part of the body morphometry analysis [34]. With advancements in deep learning and imaging post-processing techniques, it is now possible to assess both the muscle area and myosteatosis. A web-based toolkit was developed for the automatic analysis of myosteatosis on abdominal CT scans, as depicted in Fig. 3 [14]. Additionally, this toolkit enables the evaluation of visceral and subcutaneous fat areas. This toolkit utilizes deep-learning algorithms to accurately quantify and evaluate the degree of myosteatosis, thereby providing a valuable tool for comprehensive body morphometry analysis.

Fig. 3
Web-based toolkit for myosteatosis evaluation.
Sfat = subcutaneous fat area; Vfat = visceral fat area; IMA = intermuscular adipose tissue content; LAMA = low-attenuation muscle area; NAMA = normal-attenuation muscle area.

• Click for larger image
• Download as PowerPoint slide

Evaluation of the clinical impact following gastrectomy

Gastrectomy influences one of the most crucial human functions, eating, owing to early satiety, loss of appetite, dumping syndrome, reflux gastritis/esophagitis, and a variety of other adverse events [7]. Consequently, poor eating habits can cause considerable weight loss, nutritional deficiencies, and decreased physical activity. These changes can lead to sarcopenia, frailty, and mortality [6]. Furthermore, individuals undergoing subtotal and total gastrectomies exhibit substantial variations in weight change and body morphometric alterations, with remnant stomach volume (RSV) being a key contributing factor [35]. We used AI body morphometry techniques to evaluate the clinical impact of gastrectomies.

First, a volumetric measurement method for RSV was developed using postgastrectomy CT gastrography [35]. In patients who have undergone gastrectomy, CT gastrography with the oral administration of effervescent granules is performed to adequately distend the remnant stomach for volume measurements. At our institution, a three-dimensional (3D) volumetry technique has been established that generates volume-rendering images using semiautomatic segmentation software that distinguishes the luminal space of the remnant stomach containing air (Fig. 4).

Fig. 4
Combined use of three-dimensional volumetry of RSV and artificial intelligence body morphometry.
RSV = remnant stomach volume.

• Click for larger image
• Download as PowerPoint slide

By employing AI-based body morphometry techniques to measure the muscle mass and utilizing 3D volumetry for RSV, Lee et al. [6] investigated the impact of RSV on nutrition, anemia, and sarcopenia. During the first year after gastrectomy, all the patients experienced a decrease in the visceral fat area on CT, leading to significant weight loss. The extent of this decrease was influenced by the RSV, which was the highest in the total gastrectomy group without RSV and lowest in the large RSV group [6]. Kim et al. [15] reported that a larger RSV was associated with better preservation of muscle and fat after distal gastrectomy, as illustrated in Fig. 4.

These findings suggest the potential for a paradigm shift in the guidelines for gastrectomy with regard to preserving the RSV [6, 15]. Currently, gastrectomy for gastric cancer involves the removal of at least two-thirds of the stomach to minimize the risk of tumor recurrence [29]. However, based on the results of these studies, modifications to the guidelines could be considered to prioritize the preservation of the RSV and optimize postoperative outcomes, nutritional status, and body composition. Further research and clinical trials are warranted to validate these findings and support potential guideline changes for gastrectomy procedures.

Radiomics-based approaches

Radiomics approaches primarily utilize radiomic features extracted from medical images, such as CT or MRI, to predict the prognosis of patients with gastric cancer. These approaches often focus on disease-free survival with the aim of evaluating the likelihood of recurrence following gastrectomy. By analyzing and quantifying a wide range of radiomic features, radiomic models offer the potential to enhance prognostic assessments, enabling clinicians to identify patients at higher risk of recurrence and tailor treatment strategies accordingly [16, 17]. However, a notable challenge with these radiomics approaches is the need for manual segmentation of tumors, which poses a significant hurdle to their integration into routine clinical practice.

Multi-modal approach

Overall survival after gastrectomy is influenced by a multitude of factors: 1) tumor factors, including TNM staging or histology; 2) demographic factors, such as age, sex, performance status, and socioeconomic status; 3) pathophysiological factors, such as neutrophil/lymphocyte ratio, nutritional status, and body morphometric changes after gastrectomy; and 4) treatment factors, such as adjuvant chemotherapy and surgical procedures [5]. Considering the multifactorial nature of overall survival, it is essential to comprehensively assess these factors when determining the prognostic outcomes and developing personalized treatment strategies for patients who have undergone gastrectomy for gastric cancer [8].

Traditionally, statistical prognostic models, such as the Cox proportional hazards regression models or nomogram models, have been employed to predict the overall survival. However, these models were limited by the number of variables included in the model. Consequently, AI-based prediction models have garnered increasing attention, particularly owing to the advantages of deep learning in handling large clinical datasets with nonlinear effects and interactions between variables. Consequently, growing emphasis exists on multimodal AI models that integrate both CT imaging and EHRs to enhance the accuracy of prognostic models [18].

Recently, Chung et al. [8] developed an AI prognostic model for predicting 5-year survival using a substantial dataset of over 4,000 cases with 63 variables. These variables include nutrition, skeletal muscle mass, visceral/subcutaneous fat, sarcopenia, obesity, comorbidities, interval changes before and after surgery, and cancer-related variables [8]. In the external validation, they achieved promising results, with an AUC of 0.8903, sensitivity of 86.96%, and specificity of 74.60% for predicting 5-year survival. Despite these encouraging findings, the authors emphasize the need for further validation in larger populations to confirm the robustness and generalizability of the model.