Diagnostic value of artificial intelligence computer-assisted diagnosis (computer assisted-diagnosis eye function) for colorectal polyps

Abstract

BACKGROUND

The gold standard for colorectal polyp screening is currently colonoscopy, but the miss rate is still high and the adenoma detection rate and polyp detection rate are still low. The risk factors include the patient, operators, and the tools used. The use of artificial intelligence (AI) in colonoscopy has gained popularity by assisting endoscopists in the detection and characterization of polyps.

AIM

To evaluate the diagnostic performance of AI-assisted colonoscopy [computer assisted diagnosis (CAD) eye function] for colorectal polyp characterization.

METHODS

This study used a cross-sectional design conducted at the Gastrointestinal Endoscopy Center of Dr. Cipto Mangunkusumo Hospital in January-May 2024 on adult patients with suspected colorectal polyps.

RESULTS

A total of 60 patients with 100 polyps were involved in this study. Based on the results of the examination, it was found that the AI CAD eye function examination had a sensitivity of 79.17%, specificity of 75.00%, positive predictive value (PPV) of 89.06%, negative predictive value (NPV) of 58.33%, and accuracy of 78.00%. In polyps with diminutive size, sensitivity was 86.27%, specificity was 60.00%, PPV was 95.65%, NPV was 30.00%, and accuracy was 83.93%. Meanwhile, in polyps with non-diminutive size, sensitivity was 61.90%, specificity was 78.26%, PPV was 72.22%, NPV was 69.23%, and accuracy was 70.45%. In polyps on the left side of the colon, sensitivity was 78.85%, specificity was 81.25%, PPV was 93.18%, NPV was 54.17%, and accuracy was 79.41%. Meanwhile, in right-sided polyps the sensitivity was 80.00%, specificity was 66.67%, PPV was 80.00%, NPV was 66.67%, and accuracy was 75.00%. In sessile polyps the sensitivity was 81.54%, specificity was 50.00%, PPV was 91.38%, NPV was 29.41%, and accuracy was 77.33%. Meanwhile, in non-sessile polyps, the sensitivity was 57.14%, specificity was 88.89%, PPV was 66.67%, NPV was 84.21%, and accuracy was 80.00%.

CONCLUSION

AI CAD eye function examination had a high sensitivity value in diminutive, sessile polyps and right-sided polyps and a high specificity in non-diminutive, non-sessile polyps and left-sided polyps.

Key Words: Artificial intelligence; Computer-assisted diagnosis eye function; Colorectal polyps; Inflammatory bowel disease; Artificial intelligence assisted

Core Tip: Colonoscopy is currently the gold standard for screening colorectal polyps. However, the miss rate remains high, and the adenoma detection rate and polyp detection rate remain low. The risk factors included patients, operators, and the tools used. The use of artificial intelligence in colonoscopy has gained popularity because it helps endoscopists detect and characterize polyps. This cross-sectional study was conducted on adult patients with suspected colorectal polyps. The overall performance of the artificial intelligence (computer-assisted diagnosis eye function) showed a sensitivity of 79.17%, specificity of 75.00%, positive predictive value of 89.06%, negative predictive value of 58.33%, and accuracy of 78.00%.

INTRODUCTION

Colorectal cancer (CRC) is a leading cause of cancer-related mortality worldwide[1]. Accounting for 10% of all cancer cases, it is the third-leading cause of cancer-related deaths globally[2]. The prevalence of CRC varies globally, ranging from 30 per 100000 to 40 per 100000 in the United States, 15-30 per 100000 in Europe, and less than 5-10 per 100000 in South America and Asia[3]. In Indonesia CRC cases reached 34189 in 2020[4]. CRC is the third most common cancer among both sexes[5] with a lifetime risk of approximately 6% in the United States[6]. Despite advancements in treatment the 5-year survival rate remains around 55%[7].

Colonoscopy-based screening programs are crucial for reducing mortality by enabling early diagnosis and detection of precancerous lesions. Colonoscopy can reduce CRC incidence by 70%-80%[8]. However, screening at the population level involves significant financial and economic costs[9], primarily due to histopathology costs post-polypectomy despite most resections being small polyps representing over 90% of all lesions[10]. Furthermore, long-term effectiveness of colonoscopy is influenced by various factors, including operator skill and potential complications such as bleeding, perforation, and infection[11]. Miss rates for adenomas, particularly small ones, can be significant (24%-35%)[12,13], and diagnosing small polyps is challenging due to sampling difficulties and dependence on polyp location and endoscopist experience[14]. This is consistent with interval cancer incidence, which represents cancers diagnosed after screening but before surveillance (around 3%-5%)[15].

Quality metrics in colonoscopy screening include adenoma detection rate (ADR), bowel preparation rate, cecal intubation rate, and withdrawal time. Increasing ADR by 1% can reduce CRC risk by 3%[16]. Innovations such as virtual and dye-spray chromoendoscopy, along with additional devices, can enhance ADR, particularly in low detection scenarios[17]. However, these strategies are operator-dependent and require learning time, and individual experience and preferences may affect their efficacy.

Currently, colonoscopy remains the gold standard for colorectal polyp screening, yet sensitivity has not reached 100% with high miss rates and low ADR and polyp detection rate impacting outcomes. Artificial intelligence (AI) technologies offer potential improvements in polyp detection and characterization. AI, particularly deep convolutional neural networks, has shown promise in enhancing ADR and polyp detection rate[11]. AI assists endoscopists by highlighting areas requiring closer examination and differentiating polyp types, thus potentially increasing ADR[18]. AI has demonstrated promising results in real-time polyp detection and characterization, offering faster and more accurate diagnoses compared to traditional histopathology[19]. Computer-assisted diagnosis (CAD) systems for polyp detection and characterization have evolved with advancements in computer technology and deep learning algorithms[20]. By automating detection and characterization, AI addresses issues related to perception and histological prediction, improving diagnostic efficiency and patient comfort[21].

This study aimed to evaluate the diagnostic efficacy of the AI-based CAD system in colorectal polyp characterization, correlating findings with histopathological examination based on polyp size, location, and morphology.

MATERIALS AND METHODS

This cross-sectional study evaluated the diagnostic performance of AI-assisted colonoscopy (CAD eye function) for colorectal polyp characterization. The study was conducted at the Gastrointestinal Endoscopy Center of Dr. Cipto Mangunkusumo Hospital. Data were collected from medical records and colonoscopy procedures utilizing AI (CAD eye function). Sampling occurred if colorectal polyps were detected.

The target population for this study included patients aged 18 years and older undergoing colonoscopy with AI (CAD eye function). Inclusion criteria were: Patients aged 18 or older who had risk factors for colorectal polyps and who received care at Dr. Cipto Mangunkusumo Hospital from January to May 2024. Exclusion criteria were: Incomplete medical records and a Boston bowel preparation scale score of less than 6. The sample was drawn from individuals who meet these inclusion and exclusion criteria. According to the sample size formula for descriptive categorical research, a total of 57 participants was needed.

Patients with suspected colorectal polyps who visited the Gastroenterology Clinic or were admitted to Dr. Cipto Mangunkusumo Hospital were selected for the study. After providing informed consent these patients underwent a colonoscopy using AI (CAD eye function). Detected polyps were documented in terms of their location, size, and morphology. The polyps were subjected to histopathological analysis at the RSUPN Dr. Cipto Mangunkusumo Pathology Laboratory.

Statistical analysis

Data processing were conducted using Statistical Package for the Social Sciences 22.0. Clinical and participant characteristics were reported as mean ± SD for normally distributed data or median with range for non-normally distributed data. The Kolmogorov-Smirnov test was used to assess normality. Bivariate analysis involved χ² or Fisher's exact tests for categorical variables and t-test or Mann-Whitney test for numerical variables. The performance of AI-assisted colonoscopy was evaluated based on its sensitivity and specificity in detecting polyps.

RESULTS

The median age of subjects in this study was 61.5 years with an interquartile range of 20-76 years. The most common age group was 60-69 years, comprising 23 subjects (38.3%). Gender distribution was balanced with 30 females (50.0%) and 30 males (50.0%). Ethnically, most subjects were Javanese (28, 46.7%). Obesity was present in 12 subjects (20.0%) while 48 subjects (80.0%) were not obese. Inflammatory bowel disease was observed in 30.0% of subjects (18 individuals) with 70.0% (42 individuals) having no inflammatory history. Diabetes mellitus (DM) was reported in 12 subjects (20.0%), and 48 subjects (80.0%) were without DM. A family history of polyps was found in 8 subjects (13.3%) while 52 subjects (86.7%) had no such history. A family history of CRC was noted in 6 subjects (10.0%) with 54 subjects (90.0%) having no family history of CRC (Table 1).

Table 1 Basic characteristics of patients.

Variable	Total (n = 60)
Age (years), median (interquartile range)	61.5 (20-76)
Age
< 30 years	4 (6.7)
30-39 years	4 (6.7)
40-49 years	6 (10.0)
50-59 years	13 (21.7)
60-69 years	23 (38.3)
> 70 years	10 (16.7)
Gender
Female	30 (50.0)
Male	30 (50.0)
Ethnicity
Javanese	28 (46.7)
Sundanese	5 (8.3)
Minangkabau	2 (3.3)
Batak	9 (15.0)
Chinese	9 (15.0)
Other	7 (11.7)
Obesity
Yes	12 (20.0)
No	48 (80.0)
Inflammatory bowel disease
Yes	18 (30.0)
No	42 (70.0)
Diabetes mellitus
Yes	12 (20.0)
No	48 (80.0)
Family history of polyps
Yes	8 (13.3)
No	52 (86.7)
Family history of colorectal cancer
Yes	6 (10.0)
No	54 (90.0)

Data are presented as n (%).

Most polyps measured 1-5 mm in size, accounting for 56% of the cases. Larger polyps, specifically those ≥ 10 mm were less frequent (just 33%), and only 11% of polyps fell into the 6-9 mm size range. The predominant location for polyps was the rectosigmoid region with 51% of polyps found there. The next most common locations were the ascending colon (19%), descending colon (17%), transverse colon (8%), and cecum (5%). In terms of morphology sessile polyps were the most prevalent, comprising 75% of the samples, followed by pedunculated polyps at 16%, and semipedunculated polyps at 9% (Table 2).

Table 2 Basic characteristics of patients based on the size, location, and morphology of polyps.

Variable	Total (n = 100)
Polyp size
1-5 mm	56 (56)
6-9 mm	11 (11)
≥ 10 mm	33 (33)
Polyp location
Rectosigmoid polyp	51 (51)
Descending colon polyp	17
Transverse colon polyp	8 (8)
Ascending colon polyp	19 (19)
Cecum polyp	5
Morfologi polyp
Sessile polyp	75
Semipedunculated polyp	9 (9)
Pedunculated polyp	16 (16)
Flat polyp	0 (0)
Other polyp	0 (0)

Data are presented as n (%).

Based on the study results, the overall performance of AI (CAD eye function) showed a sensitivity of 79.17%, specificity of 75.00%, positive predictive value (PPV) of 89.06%, negative predictive value (NPV) of 58.33%, and accuracy of 78.00% as detailed in Table 3.

Table 3 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	57	7	64
	Neoplastic	15	21	36
	Total	72	28	100
Sensitivity (%)	79.17
Specificity (%)	75.00
Positive predictive value (%)	89.06
Negative predictive value (%)	58.33
Accuracy	78.00

For diminutive polyps, which are 1-5 mm in size, CAD eye function demonstrated a high sensitivity of 86.27%, but a lower specificity of 60.00% with a PPV of 95.65% and a NPV of 30.00%, resulting in an accuracy of 83.93% (Table 4). Conversely, for non-diminutive polyps, which are larger than 5 mm, the sensitivity was lower at 61.90%, but the specificity was higher at 78.26% with a PPV of 72.22% and an NPV of 69.23%, leading to an accuracy of 70.45% as detailed in Table 5.

Table 4 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in diminutive polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	44	2	46
	Neoplastic	7	3	10
	Total	51	5	56
Sensitivity (%)	86.27
Specificity (%)	60.00
Positive predictive value (%)	95.65
Negative predictive value (%)	30.00
Accuracy	83.93

Table 5 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in non-diminutive polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	13	5	18
	Neoplastic	8	18	26
	Total	21	23	44
Sensitivity (%)	61.90
Specificity (%)	78.26
Positive predictive value (%)	72.22
Negative predictive value (%)	69.23
Accuracy	70.45

Detecting left-sided colon polyps showed a sensitivity of 78.85%, specificity of 81.25%, PPV of 93.18%, NPV of 54.17%, and accuracy of 79.41% (Table 6). In contrast, for right-sided colon polyps, the sensitivity was 80.00%, specificity was 66.67%, PPV was 80.00%, NPV was 66.67%, and accuracy was 75.00% (Table 7).

Table 6 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in left-sided colon polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	41	3	44
	Neoplastic	11	13	24
	Total	52	16	68
Sensitivity (%)	78.85
Specificity (%)	81.25
Positive predictive value (%)	93.18
Negative predictive value (%)	54.17
Accuracy	79.41

Table 7 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in right-sided colon polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	16	4	20
	Neoplastic	4	8	12
	Total	20	12	32
Sensitivity (%)	80.00
Specificity (%)	66.67
Positive predictive value (%)	80.00
Negative predictive value (%)	66.67
Accuracy	75.00

For sessile polyps CAD eye function demonstrated a sensitivity of 81.54%, specificity of 50.00%, PPV of 91.38%, NPV of 29.41%, and accuracy of 77.33% when compared with the gold standard (histopathology) (Table 8). Conversely, for non-sessile polyps the sensitivity was lower at 57.14%, but the specificity was higher at 88.89%, with a PPV of 66.67% and an NPV of 84.21%, resulting in an accuracy of 80.00% as detailed in the Table 9.

Table 8 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in sessile polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	53	5	58
	Neoplastic	12	5	17
	Total	65	10	75
Sensitivity (%)	81.54
Specificity (%)	50.00
Positive predictive value (%)	91.38
Negative predictive value (%)	29.41
Accuracy	77.33

Table 9 Comparison of artificial intelligence (computer-assisted diagnosis eye function) findings with histopathology results in non-sessile polyps.

Examination result	Histopathology
		Hyperplastic	Neoplastic	Total
Computer-assisted diagnosis eye	Hyperplastic	4	2	6
	Neoplastic	3	16	19
	Total	7	18	25
Sensitivity (%)	57.14
Specificity (%)	88.89
Positive predictive value (%)	66.67
Negative predictive value (%)	84.21
Accuracy	80.00

DISCUSSION

Comparison of AI CAD eye function and histopathology examination results

Colonoscopy using AI CAD eye function showed good performance in diagnosing colorectal polyps. Based on the results, the sensitivity and specificity of CAD eye function in detecting hyperplastic polyps were 79.17% and 75.00%, respectively. The PPV of 89.06% and NPV of 58.33% showed that the method was quite reliable in identifying hyperplastic and neoplastic polyps. This indicates good diagnostic value. AI has advantages over histopathology in terms of diagnostic speed time. AI can determine polyp characteristics within moments while histopathology takes 1-2 weeks. Histopathology remains the gold standard as AI still has limitations.

Comparison of CAD eye function with histopathology based on polyp characteristics

Polyp size: Based on size, diminutive polyps (≤ 5 mm) had a sensitivity of 86.27%, specificity of 60.00%, PPV of 95.65%, NPV of 30.00%, and accuracy of 83.93%. Meanwhile, non-diminutive polyps (> 5 mm) showed a sensitivity of 61.90%, specificity of 78.26%, PPV of 72.22%, NPV of 69.23%, and accuracy of 70.45%. This may be due to the fact that AI (CAD eye function) observation of large polyps did not cover all polyp surfaces. Only the edges of the polyps and large polyps in the subjects of this study were mostly pedunculated polyps, which are generally mobile, making it difficult for AI to characterize perfectly. It is suggested to develop AI with more specific features for large polyps.

Polyp morphology: Non-sessile polyps had a sensitivity of 57.14%, specificity of 88.89%, PPV of 66.67%, NPV of 84.21%, and accuracy of 80.00% while sessile polyps showed a sensitivity of 81.54%, specificity of 50.00%, PPV of 91.38%, NPV of 29.41%, and accuracy of 77.33%. In this study it was found that non-sessile polyps had low sensitivity, likely due to the fact that the majority of non-sessile polyps in this study were pedunculated and semipedunculated polyps, which have a larger size and are mobile, so that the CAD eye function cannot cover the entire surface of the polyp and because it is mobile, making it difficult for AI to characterize perfectly. Sessile polyps have low specificity that may be due to the fact that AI only assesses the surface of the polyp while the neoplastic process starts from the submucosa, which is not assessed by AI.

Polyp location: On the left side of the colon, the sensitivity was recorded at 78.85%, specificity of 81.25%, PPV of 93.18%, NPV of 54.17%, and accuracy of 79.41%. Meanwhile, on the right side of the colon, the sensitivity reached 80.00%, specificity of 66.67%, PPV of 80.00%, NPV of 66.67%, and accuracy of 75.00%. This is because the anatomy and structure of the right colon have fewer folds and is more fixed, making it easier for AI to characterize polyps while the left side has many folds and is not fixed.

Limitation, strength, and future direction

The limitation of this study was that the input data used by AI needed to be varied to improve sensitivity and specificity. AI also has limitations in large and mobile object features.

With future data variations, it is expected to increase its sensitivity and specificity on large and moving objects. As of now histopathology examination is still the gold standard and reference for data input to AI.

This study was the first to thoroughly explore the diagnostic capabilities of AI on colorectal polyp characteristics. To further explore the application of AI, it is recommended that there be a cost-effective study between AI and histopathology so that it can be a consideration for policy makers in using AI in improving CRC screening.

CONCLUSION

Colonoscopy with the AI CAD eye function demonstrates overall good diagnostic value and can be used to improve the characteristics of colorectal polyps. AI CAD eye function examination has a high sensitivity value in diminutive, sessile polyps and right-sided colon and high specificity in non-diminutive, non-sessile and left-sided colon. Histopathology is still the gold standard to evaluate the result of the AI. Therefore, AI can improve the future result and enhance the sensitivity and specificity.