Preoperative determination of lymph node metastasis (LNM) risk is crucial for patients with submucosal invasive (T1) colorectal cancer (CRC) following endoscopic resection. While colorectal intramucosal cancer (Tis)/high-grade dysplasia is suitable for endoscopic resection due to the absence of LNM, surgical intervention remains the standard treatment for cancers invading beyond the muscularis propria layer (T2).¹ Notably, approximately 10% of patients with T1 CRC (positioned between Tis and T2 stages) exhibit extraintestinal LNM, leading to a pivotal decision between opting for endoscopic treatment or surgical resection.^2,3 The curative potential of endoscopic resection in T1 CRC or the necessity for additional surgical intervention is contingent upon accurately predicting the risk of LNM from the pathological diagnosis.

The prevalence of endoscopically resected T1 CRC is anticipated to rise, driven by an overall increase in CRC incidence, a higher detection rate of T1 CRC, and a growing preference for endoscopic over surgical intervention. Advances in endoscopic techniques, such as endoscopic submucosal dissection, endoscopic intermuscular dissection, per anal endoscopic myectomy, and endoscopic full-thickness resection, are expected to bolster this trend further.^4,5 Recent meta-analyses challenge the view that deep submucosal invasion (submucosal invasion ≥1,000 µm) inherently presents a high risk of LNM.⁶ Findings suggest that deep submucosal invasion in CRCs lacking additional risk factors such as lymphovascular invasion, poorly differentiated histology, and tumor budding is associated with a relatively low LNM-positive rate of 1.3% (95% confidence interval [CI], 0% to 2.4%).¹ Such a reassessment of risk may further subdivide deep submucosal invasion CRCs into low and high risk and shift the treatment paradigm towards more conservative endoscopic management for low-risk deep submucosal invasion cases.² The difficulty in preoperatively assessing critical factors linked to LNM, such as lymphovascular invasion and histological grade, often leads to a conservative approach: opting for endoscopic removal first to avoid over-treatment, with the caveat of ensuring negative vertical margins. This strategy aims to minimize surgery-related morbidity and preserve quality of life without compromising the oncologic outcome. Decision-making in T1 CRC treatment is multifaceted, and it considers patient preferences, the potential for cure, treatment invasiveness, and the cost implications. However, the cornerstone of this decision-making process is the stratification of LNM risk, underscoring the urgent need for precise and reliable predictive models.

Current guidelines specify the criteria for additional bowel resection following endoscopic resection of T1 CRC, having identified a subset of patients for whom endoscopic treatment may be curative.^1,7-10 However, there is a need for enhanced precision in stratifying the risk of LNM. According to these guidelines, LNM is present in only about 10% of cases undergoing surgery, suggesting that the vast majority of surgical interventions might be unnecessary. To tackle the limitations of current guidelines in accurately predicting LNM, we introduce three innovative predictive models.¹¹

1. Artificial neural network model

Developed using a dataset of 5,131 T1 CRCs from seven centers in Japan, collected between 1997 and 2017, this artificial intelligence (AI) model employs machine learning to evaluate metastasis risk based on eight parameters: patient sex and age, tumor size, location and morphology, lymphatic and vascular invasion, and histological type.¹² Six of these centers contributed to training the model, with one center performing external validation. The artificial neural network significantly outperformed existing guidelines, achieving an area under the curve (AUC) of 0.83, which is markedly higher than the 0.57 AUC of the current CRC treatment guidelines (p<0.001).

2. Nomogram

This model provides a visual representation of calculated predictive probabilities, clearly outlining the impact of each variable.¹³ It was developed using data from 6,105 cases across 27 centers in Japan, from 2009 to 2016. Out of these, 3,080 cases were used to develop the nomogram, and 1,593 cases were reserved for testing. The nomogram incorporated six factors: patient sex, tumor location, tumor grade, lymphovascular invasion, tumor budding, and submucosal invasion depth. It achieved a concordance statistic (C-statistic) of 0.790, surpassing the 0.777 C-statistic of current guidelines.

3. Whole slide image-based AI model

Addressing the reproducibility issues associated with pathological assessments, a new pathologist-independent model is being developed.¹⁴ This model evaluates LNM risk directly from hematoxylin-eosin stained images, eliminating the need for human diagnosis. It operates by dividing a hematoxylin-eosin stained image (×40) of a T1 CRC into 224×224 pixel patches, stratifying each patch according to ten levels of metastasis risk, and then aggregating these to calculate the overall risk for the lesion. This methodology has demonstrated high diagnostic accuracy, with an AUC of 0.72 for the whole slide image-based AI, significantly reducing the 21% of over-surgery and achieving a sensitivity of 100% and a specificity of 35%. To date, a total of four similar studies have been published.^14-17

These developments represent a significant advancement in the precision of diagnosing LNM in T1 CRC, aiming to refine treatment strategies and reduce unnecessary surgical interventions.

Establishing the diagnostic accuracy for LNM prediction in T1 CRC involves a critical balance between achieving a high enough sensitivity to detect all potential cases of LNM while maintaining enough specificity to avoid unnecessary surgeries. This balance is crucial, because missed LNM can lead to disease recurrence and death, whereas overly aggressive treatment can increase the rates of morbidity and mortality associated with surgery.

1. Sensitivity

While the goal of 100% sensitivity is laudable, this includes an inherent risk of false positives and, therefore, requires a more nuanced approach. An acceptable level of sensitivity should minimize the risk of missing LNM without significantly increasing unnecessary surgical interventions.

2. Specificity and PPV

Elevating specificity and positive predictive value (PPV) aim to reduce over-treatment and limit surgical interventions to patients needing them based on a high probability of LNM. The challenge lies in enhancing these two metrics without substantially impacting the model’s sensitivity.

3. Postoperative mortality rates

Using postoperative mortality rates as a benchmark for setting sensitivity and specificity targets offers a pragmatic solution. This approach balances the risk of missed LNM (and the potential for endoscopic treatment alone) against the morbidity and mortality associated with surgical treatments.

4. Reference points for model accuracy

One approach defining this standard involves comparing postoperative mortality rates: the risk of death from missed LNM after endoscopic treatment alone should be similar to or lower than the risk of surgery-related mortality. As noted in Table 1, in Japan, the 90-day postoperative mortality rates are 2.0% for right hemicolectomy (n=22,410) and 0.6% for low anterior resection (n=21,262), encompassing both early and advanced-stage cancers.¹⁸ Similarly, in the United States, age-specific surgery-related mortality rates were reported (total n=1,043,108) across various age groups, showing an increase in mortality with age.¹⁹ For T1 CRC, the surgery-related mortality rate was similar to that of a Dutch study, 1.7% (n=5,170), suggesting that these rates can be used as reference values for acceptable sensitivity thresholds in predictive models.²⁰ Furthermore, the Japanese Society for Cancer of the Colon and Rectum project on 2,468 cases of T1 CRC in Japan revealed a low LNM-positive rate of 0.3% (1/325; 95% CI, 0.0% to 1.7%) in the guideline-defined endoscopy curative (low-risk) group.¹ The sensitivity, specificity, and PPV achieved were 99.6% (95% CI, 98.0% to 100%), 14.8% (95% CI, 13.3% to 16.3%), and 12.6% (95% CI, 11.3% to 14.1%), respectively, indicating acceptable sensitivity.

Table 1 . Postoperative Mortality of Colorectal Cancer.

Author (year)	Country	Definition	T-stage	Postoperative mortality, %	No. of patients	Descriptions
Marubashi et al. (2021)18	Japan	90-day mortalities	All	0.6	21,262	Low anterior resection
				2.0	22,410	Right hemicolectomy
Vermeer et al. (2019)20	Netherlands	30-day mortalities	T1	1.7	5,170	-
			T2-T3	2.5	34,643
Jafari et al. (2014)19	USA	In-hospital mortalities	All	45–64 yr: 1.3	377,129	-
				65–69 yr: 2.0	132,807
				70–74 yr: 2.9	143,132
				75–79 yr: 3.7	154,433
				80–84 yr: 4.9	128,686
				≥85 yr: 8.0	106,921

Future predictive models should aim to match the high sensitivity levels outlined in current guidelines while seeking to improve specificity and PPV. This dual objective acknowledges the complexity of balancing diagnostic accuracy with the clinical imperative to do no harm. It is also necessary to provide evidence regarding postoperative mortality for T1 CRC.

This work was supported by JSPS KAKENHI (Grant number 22K16500).

No potential conflict of interest relevant to this article was reported.

Study conception: K.I., S.K., H.M. Acquisition data: K.I. Interpretation of data: K.I., M.M., Y.K., Y.T. Drafting of the article: K.I. Critical revision of the article for important intellectual content: S.K., M.M., K.G.Y., T.N., Y.K., Y.T., H.M. Final approval of the article: all authors.