Scholarlink

Objectives <p>Computed tomography (CT) colonography is increasingly recognized as a valuable modality for diagnosing colorectal lesions, however, the interpretation workload remains challenging for physicians. Deep learning-based artificial intelligence (AI) algorithms have been employed for imaging diagnoses. In this study, we examined the sensitivity of neoplastic lesions in CT colonography images.</p> Methods <p>Lesion location and size were evaluated during colonoscopy and a large-scale database including a dataset for AI learning and external validation was created. The DICOM data used as training data and internal validation data (total 453 patients) for this study were colorectal cancer screening test data from two multicenter joint trial conducted in Japan and data from two institutions. External validation data (137 patients) were from other two institutions. Lesions were categorized into ≥6 mm, 6 to 10 mm, and ≥10 mm. During this study, we adopted a neural network structure that was designed based on the faster R-CNNs to detect colorectal lesion. The sensitivity of detecting colorectal lesions was verified when one and two positions were integrated.</p> Results <p>Internal validation yielded sensitivity of 0.815, 0.738, and 0.883 for lesions ≥6 mm, 6 to 10 mm, and ≥10 mm, respectively, with a false lesion limit of three. Two external validation produced rates of 0.705 and 0.707, 0.575 and 0.573, and 0.760 and 0.779 for each lesion category. Combining two positions for each patient in calculating the sensitivity resulted in significantly improved rates for each lesion category.</p> Conclusions <p>The sensitivity of CT colonography images using the AI algorithm was improved by integrating evaluations in two positions. Validation experiments involving radiologists who can interpret images as well as AI to determine the auxiliary diagnosis can reduce the workload of physicians.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development and validation of computer-aided detection for colorectal neoplasms using deep learning incorporated with computed tomography colonography

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月8日

查看全文
  • Shungo Endo,
  • Koichi Nagata,
  • Kenichi Utano,
  • Satoshi Nozu,
  • Takaaki Yasuda,
  • Ken Takabayashi,
  • Michiaki Hirayama,
  • Kazutomo Togashi,
  • Hiromasa Ohira

摘要

Objectives

Computed tomography (CT) colonography is increasingly recognized as a valuable modality for diagnosing colorectal lesions, however, the interpretation workload remains challenging for physicians. Deep learning-based artificial intelligence (AI) algorithms have been employed for imaging diagnoses. In this study, we examined the sensitivity of neoplastic lesions in CT colonography images.

Methods

Lesion location and size were evaluated during colonoscopy and a large-scale database including a dataset for AI learning and external validation was created. The DICOM data used as training data and internal validation data (total 453 patients) for this study were colorectal cancer screening test data from two multicenter joint trial conducted in Japan and data from two institutions. External validation data (137 patients) were from other two institutions. Lesions were categorized into ≥6 mm, 6 to 10 mm, and ≥10 mm. During this study, we adopted a neural network structure that was designed based on the faster R-CNNs to detect colorectal lesion. The sensitivity of detecting colorectal lesions was verified when one and two positions were integrated.

Results

Internal validation yielded sensitivity of 0.815, 0.738, and 0.883 for lesions ≥6 mm, 6 to 10 mm, and ≥10 mm, respectively, with a false lesion limit of three. Two external validation produced rates of 0.705 and 0.707, 0.575 and 0.573, and 0.760 and 0.779 for each lesion category. Combining two positions for each patient in calculating the sensitivity resulted in significantly improved rates for each lesion category.

Conclusions

The sensitivity of CT colonography images using the AI algorithm was improved by integrating evaluations in two positions. Validation experiments involving radiologists who can interpret images as well as AI to determine the auxiliary diagnosis can reduce the workload of physicians.