A Semi-Supervised Learning Framework for Classifying Colorectal Neoplasia Based on the NICE Classification

Labelling medical images is an arduous and costly task that necessitates clinical expertise and large numbers of qualified images. Insufficient samples can lead to underfitting during training and poor performance of supervised learning models. In this study, we aim to develop a SimCLR-based semi-supervised learning framework to classify colorectal neoplasia based on the NICE classification. First, the proposed framework was trained under self-supervised learning using a large unlabelled dataset; subsequently, it was fine-tuned on a limited labelled dataset based on the NICE classification. The model was evaluated on an independent dataset and compared with models based on supervised transfer learning and endoscopists using accuracy, Matthew's correlation coefficient (MCC), and Cohen's kappa. Finally, Grad-CAM and t-SNE were applied to visualize the models' interpretations. A ResNet-backboned SimCLR model (accuracy of 0.908, MCC of 0.862, and Cohen's kappa of 0.896) outperformed supervised transfer learning-based models (means: 0.803, 0.698, and 0.742) and junior endoscopists (0.816, 0.724, and 0.863), while performing only slightly worse than senior endoscopists (0.916, 0.875, and 0.944). Moreover, t-SNE showed a better clustering of ternary samples through self-supervised learning in SimCLR than through supervised transfer learning. Compared with traditional supervised learning, semi-supervised learning enables deep learning models to achieve improved performance with limited labelled endoscopic images.

Development of multi-class computer-aided diagnostic systems using the NICE/JNET classifications for colorectal lesions

BACKGROUND AND AIM

Diagnostic support using artificial intelligence may contribute to the equalization of endoscopic diagnosis of colorectal lesions. We developed computer-aided diagnosis (CADx) support system for diagnosing colorectal lesions using the NBI International Colorectal Endoscopic (NICE) classification and the Japan NBI Expert Team (JNET) classification.

METHODS

Using Residual Network as the classifier and NBI images as training images, we developed a CADx based on the NICE classification (CADx-N) and a CADx based on the JNET classification (CADx-J). For validation, 480 non-magnifying and magnifying NBI images were used for the CADx-N and 320 magnifying NBI images were used for the CADx-J. The diagnostic performance of the CADx-N was evaluated using the magnification rate.

RESULTS

The accuracy of the CADx-N for Types 1, 2, and 3 was 97.5%, 91.2%, and 93.8%, respectively. The diagnostic performance for each magnification level was good (no statistically significant difference). The sensitivity, specificity, positive predictive value, negative predictive value, and accuracy of the CADx-J were 100%, 96.3%, 82.8%, 100%, and 96.9% for Type 1; 80.3%, 93.7%, 94.1%, 79.2%, and 86.3% for Type 2A; 80.4%, 84.7%, 46.8%, 96.3%, and 84.1% for Type 2B; and 62.5%, 99.6%, 96.8%, 93.8%, and 94.1% for Type 3, respectively.

CONCLUSIONS

The multi-class CADx systems had good diagnostic performance with both the NICE and JNET classifications and may aid in educating non-expert endoscopists and assist in diagnosing colorectal lesions.

Magnifying Narrow Band Imaging (NBI) for the Diagnosis of Localized Colorectal Lesions Using the Japan NBI Expert Team (JNET) Classification

OBJECTIVE

The Japan NBI Expert Team (JNET) proposed a new narrow band imaging (NBI) classification system for colorectal tumors in June 2014. In this classification system, types 1, 2A, 2B, and 3 correspond to hyperplastic polyps (HPs) including sessile serrated polyps (SSPs), low-grade dysplasia (LGD), high-grade dysplasia (HGD) to shallow submucosal invasive (SM-s) carcinomas, and deep submucosal invasive (SM-d) carcinomas, respectively.

METHODS

To validate this system, we performed a retrospective image evaluation study, in which 199 colorectal tumors previously assessed by NBI magnifying endoscopy were classified by 3 blinded experienced colonoscopists using the JNET system. The results were compared with the final pathological diagnoses to determine the JNET classification's accuracy. The interobserver agreement was calculated, and the intraobserver agreement was assessed after 6 months.

RESULTS

The final pathological diagnoses identified 14 HPs/SSPs, 127 LGDs, 22 HGDs, 19 SM-s carcinomas, and 17 SM-d carcinomas. The respective sensitivities, specificities, positive predictive value, negative predictive value, and accuracies were as follows: Type 1, 85.7, 99.5, 92.3, 98.9, and 98.5%; Type 2A, 96.0, 81.9, 90.3, 92.1, and 90.9%; Type 2B, 75.6%, 90.5, 67.3, 93.4, and 87.4%; and Type 3, 29.4%, 100, 100, 93.8, and 94.0%. The interobserver agreement and the intraobserver agreement were moderate (κ value: 0.52) and excellent (κ value: 0.88), respectively. Lesions presenting as Type 2B during NBI comprised a range of colorectal tumors, including HGDs, SM-s, and SM-d.

CONCLUSIONS

The JNET classification was useful for the diagnosis of HPs/SSPs, LGDs, and SM-d, but not SM-s lesions. For low-confidence cases, magnified chromoendoscopy is recommended to ensure correct diagnoses.